diff --git a/sys/kern/vfs_bio.c b/sys/kern/vfs_bio.c index 02b13cf3b9ae..6b284c830446 100644 --- a/sys/kern/vfs_bio.c +++ b/sys/kern/vfs_bio.c @@ -1,3390 +1,3389 @@ /* * Copyright (c) 1994,1997 John S. Dyson * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice immediately at the beginning of the file, without modification, * this list of conditions, and the following disclaimer. * 2. Absolutely no warranty of function or purpose is made by the author * John S. Dyson. * * $FreeBSD$ */ /* * this file contains a new buffer I/O scheme implementing a coherent * VM object and buffer cache scheme. Pains have been taken to make * sure that the performance degradation associated with schemes such * as this is not realized. * * Author: John S. Dyson * Significant help during the development and debugging phases * had been provided by David Greenman, also of the FreeBSD core team. * * see man buf(9) for more info. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); struct bio_ops bioops; /* I/O operation notification */ struct buf_ops buf_ops_bio = { "buf_ops_bio", bwrite }; /* * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. */ struct buf *buf; /* buffer header pool */ struct mtx buftimelock; /* Interlock on setting prio and timo */ static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to); static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to); static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m); static void vfs_clean_pages(struct buf * bp); static void vfs_setdirty(struct buf *bp); static void vfs_vmio_release(struct buf *bp); static void vfs_backgroundwritedone(struct buf *bp); static int flushbufqueues(void); static void buf_daemon(void); int vmiodirenable = TRUE; SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, "Use the VM system for directory writes"); int runningbufspace; SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, "Amount of presently outstanding async buffer io"); static int bufspace; SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, "KVA memory used for bufs"); static int maxbufspace; SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, "Maximum allowed value of bufspace (including buf_daemon)"); static int bufmallocspace; SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, "Amount of malloced memory for buffers"); static int maxbufmallocspace; SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, "Maximum amount of malloced memory for buffers"); static int lobufspace; SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, "Minimum amount of buffers we want to have"); static int hibufspace; SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, "Maximum allowed value of bufspace (excluding buf_daemon)"); static int bufreusecnt; SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, "Number of times we have reused a buffer"); static int buffreekvacnt; SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, "Number of times we have freed the KVA space from some buffer"); static int bufdefragcnt; SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, "Number of times we have had to repeat buffer allocation to defragment"); static int lorunningspace; SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, "Minimum preferred space used for in-progress I/O"); static int hirunningspace; SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, "Maximum amount of space to use for in-progress I/O"); static int numdirtybuffers; SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, "Number of buffers that are dirty (has unwritten changes) at the moment"); static int lodirtybuffers; SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, "How many buffers we want to have free before bufdaemon can sleep"); static int hidirtybuffers; SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, "When the number of dirty buffers is considered severe"); static int numfreebuffers; SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, "Number of free buffers"); static int lofreebuffers; SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, "XXX Unused"); static int hifreebuffers; SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, "XXX Complicatedly unused"); static int getnewbufcalls; SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, "Number of calls to getnewbuf"); static int getnewbufrestarts; SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, "Number of times getnewbuf has had to restart a buffer aquisition"); static int dobkgrdwrite = 1; SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, "Do background writes (honoring the BX_BKGRDWRITE flag)?"); /* * Wakeup point for bufdaemon, as well as indicator of whether it is already * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it * is idling. */ static int bd_request; /* * bogus page -- for I/O to/from partially complete buffers * this is a temporary solution to the problem, but it is not * really that bad. it would be better to split the buffer * for input in the case of buffers partially already in memory, * but the code is intricate enough already. */ vm_page_t bogus_page; /* * Offset for bogus_page. * XXX bogus_offset should be local to bufinit */ static vm_offset_t bogus_offset; /* * Synchronization (sleep/wakeup) variable for active buffer space requests. * Set when wait starts, cleared prior to wakeup(). * Used in runningbufwakeup() and waitrunningbufspace(). */ static int runningbufreq; /* * Synchronization (sleep/wakeup) variable for buffer requests. * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done * by and/or. * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), * getnewbuf(), and getblk(). */ static int needsbuffer; /* * Mask for index into the buffer hash table, which needs to be power of 2 in * size. Set in kern_vfs_bio_buffer_alloc. */ static int bufhashmask; /* * Hash table for all buffers, with a linked list hanging from each table * entry. Set in kern_vfs_bio_buffer_alloc, initialized in buf_init. */ static LIST_HEAD(bufhashhdr, buf) *bufhashtbl; /* * Somewhere to store buffers when they are not in another list, to always * have them in a list (and thus being able to use the same set of operations * on them.) */ static struct bufhashhdr invalhash; /* * Definitions for the buffer free lists. */ #define BUFFER_QUEUES 6 /* number of free buffer queues */ #define QUEUE_NONE 0 /* on no queue */ #define QUEUE_LOCKED 1 /* locked buffers */ #define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ #define QUEUE_DIRTY 3 /* B_DELWRI buffers */ #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ #define QUEUE_EMPTY 5 /* empty buffer headers */ /* Queues for free buffers with various properties */ static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; /* * Single global constant for BUF_WMESG, to avoid getting multiple references. * buf_wmesg is referred from macros. */ const char *buf_wmesg = BUF_WMESG; #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ /* * Buffer hash table code. Note that the logical block scans linearly, which * gives us some L1 cache locality. */ static __inline struct bufhashhdr * bufhash(struct vnode *vnp, daddr_t bn) { return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); } /* * numdirtywakeup: * * If someone is blocked due to there being too many dirty buffers, * and numdirtybuffers is now reasonable, wake them up. */ static __inline void numdirtywakeup(int level) { if (numdirtybuffers <= level) { if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; wakeup(&needsbuffer); } } } /* * bufspacewakeup: * * Called when buffer space is potentially available for recovery. * getnewbuf() will block on this flag when it is unable to free * sufficient buffer space. Buffer space becomes recoverable when * bp's get placed back in the queues. */ static __inline void bufspacewakeup(void) { /* * If someone is waiting for BUF space, wake them up. Even * though we haven't freed the kva space yet, the waiting * process will be able to now. */ if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; wakeup(&needsbuffer); } } /* * runningbufwakeup() - in-progress I/O accounting. * */ static __inline void runningbufwakeup(struct buf *bp) { if (bp->b_runningbufspace) { runningbufspace -= bp->b_runningbufspace; bp->b_runningbufspace = 0; if (runningbufreq && runningbufspace <= lorunningspace) { runningbufreq = 0; wakeup(&runningbufreq); } } } /* * bufcountwakeup: * * Called when a buffer has been added to one of the free queues to * account for the buffer and to wakeup anyone waiting for free buffers. * This typically occurs when large amounts of metadata are being handled * by the buffer cache ( else buffer space runs out first, usually ). */ static __inline void bufcountwakeup(void) { ++numfreebuffers; if (needsbuffer) { needsbuffer &= ~VFS_BIO_NEED_ANY; if (numfreebuffers >= hifreebuffers) needsbuffer &= ~VFS_BIO_NEED_FREE; wakeup(&needsbuffer); } } /* * waitrunningbufspace() * * runningbufspace is a measure of the amount of I/O currently * running. This routine is used in async-write situations to * prevent creating huge backups of pending writes to a device. * Only asynchronous writes are governed by this function. * * Reads will adjust runningbufspace, but will not block based on it. * The read load has a side effect of reducing the allowed write load. * * This does NOT turn an async write into a sync write. It waits * for earlier writes to complete and generally returns before the * caller's write has reached the device. */ static __inline void waitrunningbufspace(void) { /* * XXX race against wakeup interrupt, currently * protected by Giant. FIXME! */ while (runningbufspace > hirunningspace) { ++runningbufreq; tsleep(&runningbufreq, PVM, "wdrain", 0); } } /* * vfs_buf_test_cache: * * Called when a buffer is extended. This function clears the B_CACHE * bit if the newly extended portion of the buffer does not contain * valid data. */ static __inline__ void vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, vm_page_t m) { GIANT_REQUIRED; if (bp->b_flags & B_CACHE) { int base = (foff + off) & PAGE_MASK; if (vm_page_is_valid(m, base, size) == 0) bp->b_flags &= ~B_CACHE; } } /* Wake up the buffer deamon if necessary */ static __inline__ void bd_wakeup(int dirtybuflevel) { if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { bd_request = 1; wakeup(&bd_request); } } /* * bd_speedup - speedup the buffer cache flushing code */ static __inline__ void bd_speedup(void) { bd_wakeup(1); } /* * Calculating buffer cache scaling values and reserve space for buffer * headers. This is called during low level kernel initialization and * may be called more then once. We CANNOT write to the memory area * being reserved at this time. */ caddr_t kern_vfs_bio_buffer_alloc(caddr_t v, int physmem_est) { /* * physmem_est is in pages. Convert it to kilobytes (assumes * PAGE_SIZE is >= 1K) */ physmem_est = physmem_est * (PAGE_SIZE / 1024); /* * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. * For the first 64MB of ram nominally allocate sufficient buffers to * cover 1/4 of our ram. Beyond the first 64MB allocate additional * buffers to cover 1/20 of our ram over 64MB. When auto-sizing * the buffer cache we limit the eventual kva reservation to * maxbcache bytes. * * factor represents the 1/4 x ram conversion. */ if (nbuf == 0) { int factor = 4 * BKVASIZE / 1024; nbuf = 50; if (physmem_est > 4096) nbuf += min((physmem_est - 4096) / factor, 65536 / factor); if (physmem_est > 65536) nbuf += (physmem_est - 65536) * 2 / (factor * 5); if (maxbcache && nbuf > maxbcache / BKVASIZE) nbuf = maxbcache / BKVASIZE; } #if 0 /* * Do not allow the buffer_map to be more then 1/2 the size of the * kernel_map. */ if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / (BKVASIZE * 2)) { nbuf = (kernel_map->max_offset - kernel_map->min_offset) / (BKVASIZE * 2); printf("Warning: nbufs capped at %d\n", nbuf); } #endif /* * swbufs are used as temporary holders for I/O, such as paging I/O. * We have no less then 16 and no more then 256. */ nswbuf = max(min(nbuf/4, 256), 16); /* * Reserve space for the buffer cache buffers */ swbuf = (void *)v; v = (caddr_t)(swbuf + nswbuf); buf = (void *)v; v = (caddr_t)(buf + nbuf); /* * Calculate the hash table size and reserve space */ for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) ; bufhashtbl = (void *)v; v = (caddr_t)(bufhashtbl + bufhashmask); --bufhashmask; return(v); } /* Initialize the buffer subsystem. Called before use of any buffers. */ void bufinit(void) { struct buf *bp; int i; GIANT_REQUIRED; LIST_INIT(&invalhash); mtx_init(&buftimelock, "buftime lock", NULL, MTX_DEF); for (i = 0; i <= bufhashmask; i++) LIST_INIT(&bufhashtbl[i]); /* next, make a null set of free lists */ for (i = 0; i < BUFFER_QUEUES; i++) TAILQ_INIT(&bufqueues[i]); /* finally, initialize each buffer header and stick on empty q */ for (i = 0; i < nbuf; i++) { bp = &buf[i]; bzero(bp, sizeof *bp); bp->b_flags = B_INVAL; /* we're just an empty header */ bp->b_dev = NODEV; bp->b_rcred = NOCRED; bp->b_wcred = NOCRED; bp->b_qindex = QUEUE_EMPTY; bp->b_xflags = 0; LIST_INIT(&bp->b_dep); BUF_LOCKINIT(bp); TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); LIST_INSERT_HEAD(&invalhash, bp, b_hash); } /* * maxbufspace is the absolute maximum amount of buffer space we are * allowed to reserve in KVM and in real terms. The absolute maximum * is nominally used by buf_daemon. hibufspace is the nominal maximum * used by most other processes. The differential is required to * ensure that buf_daemon is able to run when other processes might * be blocked waiting for buffer space. * * maxbufspace is based on BKVASIZE. Allocating buffers larger then * this may result in KVM fragmentation which is not handled optimally * by the system. */ maxbufspace = nbuf * BKVASIZE; hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); lobufspace = hibufspace - MAXBSIZE; lorunningspace = 512 * 1024; hirunningspace = 1024 * 1024; /* * Limit the amount of malloc memory since it is wired permanently into * the kernel space. Even though this is accounted for in the buffer * allocation, we don't want the malloced region to grow uncontrolled. * The malloc scheme improves memory utilization significantly on average * (small) directories. */ maxbufmallocspace = hibufspace / 20; /* * Reduce the chance of a deadlock occuring by limiting the number * of delayed-write dirty buffers we allow to stack up. */ hidirtybuffers = nbuf / 4 + 20; numdirtybuffers = 0; /* * To support extreme low-memory systems, make sure hidirtybuffers cannot * eat up all available buffer space. This occurs when our minimum cannot * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming * BKVASIZE'd (8K) buffers. */ while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { hidirtybuffers >>= 1; } lodirtybuffers = hidirtybuffers / 2; /* * Try to keep the number of free buffers in the specified range, * and give special processes (e.g. like buf_daemon) access to an * emergency reserve. */ lofreebuffers = nbuf / 18 + 5; hifreebuffers = 2 * lofreebuffers; numfreebuffers = nbuf; /* * Maximum number of async ops initiated per buf_daemon loop. This is * somewhat of a hack at the moment, we really need to limit ourselves * based on the number of bytes of I/O in-transit that were initiated * from buf_daemon. */ bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); bogus_page = vm_page_alloc(kernel_object, ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), VM_ALLOC_NORMAL); cnt.v_wire_count++; } /* * bfreekva() - free the kva allocation for a buffer. * * Must be called at splbio() or higher as this is the only locking for * buffer_map. * * Since this call frees up buffer space, we call bufspacewakeup(). */ static void bfreekva(struct buf * bp) { GIANT_REQUIRED; if (bp->b_kvasize) { ++buffreekvacnt; bufspace -= bp->b_kvasize; vm_map_delete(buffer_map, (vm_offset_t) bp->b_kvabase, (vm_offset_t) bp->b_kvabase + bp->b_kvasize ); bp->b_kvasize = 0; bufspacewakeup(); } } /* * bremfree: * * Remove the buffer from the appropriate free list. */ void bremfree(struct buf * bp) { int s = splbio(); int old_qindex = bp->b_qindex; GIANT_REQUIRED; if (bp->b_qindex != QUEUE_NONE) { KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); bp->b_qindex = QUEUE_NONE; } else { if (BUF_REFCNT(bp) <= 1) panic("bremfree: removing a buffer not on a queue"); } /* * Fixup numfreebuffers count. If the buffer is invalid or not * delayed-write, and it was on the EMPTY, LRU, or AGE queues, * the buffer was free and we must decrement numfreebuffers. */ if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { switch(old_qindex) { case QUEUE_DIRTY: case QUEUE_CLEAN: case QUEUE_EMPTY: case QUEUE_EMPTYKVA: --numfreebuffers; break; default: break; } } splx(s); } /* * Get a buffer with the specified data. Look in the cache first. We * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE * is set, the buffer is valid and we do not have to do anything ( see * getblk() ). This is really just a special case of breadn(). */ int bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, struct buf ** bpp) { return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); } /* * Operates like bread, but also starts asynchronous I/O on * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior * to initiating I/O . If B_CACHE is set, the buffer is valid * and we do not have to do anything. */ int breadn(struct vnode * vp, daddr_t blkno, int size, daddr_t * rablkno, int *rabsize, int cnt, struct ucred * cred, struct buf ** bpp) { struct buf *bp, *rabp; int i; int rv = 0, readwait = 0; *bpp = bp = getblk(vp, blkno, size, 0, 0); /* if not found in cache, do some I/O */ if ((bp->b_flags & B_CACHE) == 0) { if (curthread != PCPU_GET(idlethread)) curthread->td_proc->p_stats->p_ru.ru_inblock++; bp->b_iocmd = BIO_READ; bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; if (bp->b_rcred == NOCRED && cred != NOCRED) bp->b_rcred = crhold(cred); vfs_busy_pages(bp, 0); VOP_STRATEGY(vp, bp); ++readwait; } for (i = 0; i < cnt; i++, rablkno++, rabsize++) { if (inmem(vp, *rablkno)) continue; rabp = getblk(vp, *rablkno, *rabsize, 0, 0); if ((rabp->b_flags & B_CACHE) == 0) { if (curthread != PCPU_GET(idlethread)) curthread->td_proc->p_stats->p_ru.ru_inblock++; rabp->b_flags |= B_ASYNC; rabp->b_flags &= ~B_INVAL; rabp->b_ioflags &= ~BIO_ERROR; rabp->b_iocmd = BIO_READ; if (rabp->b_rcred == NOCRED && cred != NOCRED) rabp->b_rcred = crhold(cred); vfs_busy_pages(rabp, 0); BUF_KERNPROC(rabp); VOP_STRATEGY(vp, rabp); } else { brelse(rabp); } } if (readwait) { rv = bufwait(bp); } return (rv); } /* * Write, release buffer on completion. (Done by iodone * if async). Do not bother writing anything if the buffer * is invalid. * * Note that we set B_CACHE here, indicating that buffer is * fully valid and thus cacheable. This is true even of NFS * now so we set it generally. This could be set either here * or in biodone() since the I/O is synchronous. We put it * here. */ int bwrite(struct buf * bp) { int oldflags, s; struct buf *newbp; if (bp->b_flags & B_INVAL) { brelse(bp); return (0); } oldflags = bp->b_flags; if (BUF_REFCNT(bp) == 0) panic("bwrite: buffer is not busy???"); s = splbio(); /* * If a background write is already in progress, delay * writing this block if it is asynchronous. Otherwise * wait for the background write to complete. */ if (bp->b_xflags & BX_BKGRDINPROG) { if (bp->b_flags & B_ASYNC) { splx(s); bdwrite(bp); return (0); } bp->b_xflags |= BX_BKGRDWAIT; tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); if (bp->b_xflags & BX_BKGRDINPROG) panic("bwrite: still writing"); } /* Mark the buffer clean */ bundirty(bp); /* * If this buffer is marked for background writing and we * do not have to wait for it, make a copy and write the * copy so as to leave this buffer ready for further use. * * This optimization eats a lot of memory. If we have a page * or buffer shortfall we can't do it. */ if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && (bp->b_flags & B_ASYNC) && !vm_page_count_severe() && !buf_dirty_count_severe()) { if (bp->b_iodone != NULL) { printf("bp->b_iodone = %p\n", bp->b_iodone); panic("bwrite: need chained iodone"); } /* get a new block */ newbp = geteblk(bp->b_bufsize); /* set it to be identical to the old block */ memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); bgetvp(bp->b_vp, newbp); newbp->b_lblkno = bp->b_lblkno; newbp->b_blkno = bp->b_blkno; newbp->b_offset = bp->b_offset; newbp->b_iodone = vfs_backgroundwritedone; newbp->b_flags |= B_ASYNC; newbp->b_flags &= ~B_INVAL; /* move over the dependencies */ if (LIST_FIRST(&bp->b_dep) != NULL) buf_movedeps(bp, newbp); /* * Initiate write on the copy, release the original to * the B_LOCKED queue so that it cannot go away until * the background write completes. If not locked it could go * away and then be reconstituted while it was being written. * If the reconstituted buffer were written, we could end up * with two background copies being written at the same time. */ bp->b_xflags |= BX_BKGRDINPROG; bp->b_flags |= B_LOCKED; bqrelse(bp); bp = newbp; } bp->b_flags &= ~B_DONE; bp->b_ioflags &= ~BIO_ERROR; bp->b_flags |= B_WRITEINPROG | B_CACHE; bp->b_iocmd = BIO_WRITE; bp->b_vp->v_numoutput++; vfs_busy_pages(bp, 1); /* * Normal bwrites pipeline writes */ bp->b_runningbufspace = bp->b_bufsize; runningbufspace += bp->b_runningbufspace; if (curthread != PCPU_GET(idlethread)) curthread->td_proc->p_stats->p_ru.ru_oublock++; splx(s); if (oldflags & B_ASYNC) BUF_KERNPROC(bp); BUF_STRATEGY(bp); if ((oldflags & B_ASYNC) == 0) { int rtval = bufwait(bp); brelse(bp); return (rtval); } else if ((oldflags & B_NOWDRAIN) == 0) { /* * don't allow the async write to saturate the I/O * system. Deadlocks can occur only if a device strategy * routine (like in MD) turns around and issues another * high-level write, in which case B_NOWDRAIN is expected * to be set. Otherwise we will not deadlock here because * we are blocking waiting for I/O that is already in-progress * to complete. */ waitrunningbufspace(); } return (0); } /* * Complete a background write started from bwrite. */ static void vfs_backgroundwritedone(bp) struct buf *bp; { struct buf *origbp; /* * Find the original buffer that we are writing. */ if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) panic("backgroundwritedone: lost buffer"); /* * Process dependencies then return any unfinished ones. */ if (LIST_FIRST(&bp->b_dep) != NULL) buf_complete(bp); if (LIST_FIRST(&bp->b_dep) != NULL) buf_movedeps(bp, origbp); /* * Clear the BX_BKGRDINPROG flag in the original buffer * and awaken it if it is waiting for the write to complete. * If BX_BKGRDINPROG is not set in the original buffer it must * have been released and re-instantiated - which is not legal. */ KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2")); origbp->b_xflags &= ~BX_BKGRDINPROG; if (origbp->b_xflags & BX_BKGRDWAIT) { origbp->b_xflags &= ~BX_BKGRDWAIT; wakeup(&origbp->b_xflags); } /* * Clear the B_LOCKED flag and remove it from the locked * queue if it currently resides there. */ origbp->b_flags &= ~B_LOCKED; if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { bremfree(origbp); bqrelse(origbp); } /* * This buffer is marked B_NOCACHE, so when it is released * by biodone, it will be tossed. We mark it with BIO_READ * to avoid biodone doing a second vwakeup. */ bp->b_flags |= B_NOCACHE; bp->b_iocmd = BIO_READ; bp->b_flags &= ~(B_CACHE | B_DONE); bp->b_iodone = 0; bufdone(bp); } /* * Delayed write. (Buffer is marked dirty). Do not bother writing * anything if the buffer is marked invalid. * * Note that since the buffer must be completely valid, we can safely * set B_CACHE. In fact, we have to set B_CACHE here rather then in * biodone() in order to prevent getblk from writing the buffer * out synchronously. */ void bdwrite(struct buf * bp) { GIANT_REQUIRED; if (BUF_REFCNT(bp) == 0) panic("bdwrite: buffer is not busy"); if (bp->b_flags & B_INVAL) { brelse(bp); return; } bdirty(bp); /* * Set B_CACHE, indicating that the buffer is fully valid. This is * true even of NFS now. */ bp->b_flags |= B_CACHE; /* * This bmap keeps the system from needing to do the bmap later, * perhaps when the system is attempting to do a sync. Since it * is likely that the indirect block -- or whatever other datastructure * that the filesystem needs is still in memory now, it is a good * thing to do this. Note also, that if the pageout daemon is * requesting a sync -- there might not be enough memory to do * the bmap then... So, this is important to do. */ if (bp->b_lblkno == bp->b_blkno) { VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); } /* * Set the *dirty* buffer range based upon the VM system dirty pages. */ vfs_setdirty(bp); /* * We need to do this here to satisfy the vnode_pager and the * pageout daemon, so that it thinks that the pages have been * "cleaned". Note that since the pages are in a delayed write * buffer -- the VFS layer "will" see that the pages get written * out on the next sync, or perhaps the cluster will be completed. */ vfs_clean_pages(bp); bqrelse(bp); /* * Wakeup the buffer flushing daemon if we have a lot of dirty * buffers (midpoint between our recovery point and our stall * point). */ bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); /* * note: we cannot initiate I/O from a bdwrite even if we wanted to, * due to the softdep code. */ } /* * bdirty: * * Turn buffer into delayed write request. We must clear BIO_READ and * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to * itself to properly update it in the dirty/clean lists. We mark it * B_DONE to ensure that any asynchronization of the buffer properly * clears B_DONE ( else a panic will occur later ). * * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() * should only be called if the buffer is known-good. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. * * Must be called at splbio(). * The buffer must be on QUEUE_NONE. */ void bdirty(bp) struct buf *bp; { KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); bp->b_flags &= ~(B_RELBUF); bp->b_iocmd = BIO_WRITE; if ((bp->b_flags & B_DELWRI) == 0) { bp->b_flags |= B_DONE | B_DELWRI; reassignbuf(bp, bp->b_vp); ++numdirtybuffers; bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); } } /* * bundirty: * * Clear B_DELWRI for buffer. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. * * Must be called at splbio(). * The buffer must be on QUEUE_NONE. */ void bundirty(bp) struct buf *bp; { KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); if (bp->b_flags & B_DELWRI) { bp->b_flags &= ~B_DELWRI; reassignbuf(bp, bp->b_vp); --numdirtybuffers; numdirtywakeup(lodirtybuffers); } /* * Since it is now being written, we can clear its deferred write flag. */ bp->b_flags &= ~B_DEFERRED; } /* * bawrite: * * Asynchronous write. Start output on a buffer, but do not wait for * it to complete. The buffer is released when the output completes. * * bwrite() ( or the VOP routine anyway ) is responsible for handling * B_INVAL buffers. Not us. */ void bawrite(struct buf * bp) { bp->b_flags |= B_ASYNC; (void) BUF_WRITE(bp); } /* * bwillwrite: * * Called prior to the locking of any vnodes when we are expecting to * write. We do not want to starve the buffer cache with too many * dirty buffers so we block here. By blocking prior to the locking * of any vnodes we attempt to avoid the situation where a locked vnode * prevents the various system daemons from flushing related buffers. */ void bwillwrite(void) { if (numdirtybuffers >= hidirtybuffers) { int s; mtx_lock(&Giant); s = splbio(); while (numdirtybuffers >= hidirtybuffers) { bd_wakeup(1); needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); } splx(s); mtx_unlock(&Giant); } } /* * Return true if we have too many dirty buffers. */ int buf_dirty_count_severe(void) { return(numdirtybuffers >= hidirtybuffers); } /* * brelse: * * Release a busy buffer and, if requested, free its resources. The * buffer will be stashed in the appropriate bufqueue[] allowing it * to be accessed later as a cache entity or reused for other purposes. */ void brelse(struct buf * bp) { int s; GIANT_REQUIRED; KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); s = splbio(); if (bp->b_flags & B_LOCKED) bp->b_ioflags &= ~BIO_ERROR; if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && !(bp->b_flags & B_INVAL)) { /* * Failed write, redirty. Must clear BIO_ERROR to prevent * pages from being scrapped. If B_INVAL is set then * this case is not run and the next case is run to * destroy the buffer. B_INVAL can occur if the buffer * is outside the range supported by the underlying device. */ bp->b_ioflags &= ~BIO_ERROR; bdirty(bp); } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || (bp->b_ioflags & BIO_ERROR) || bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { /* * Either a failed I/O or we were asked to free or not * cache the buffer. */ bp->b_flags |= B_INVAL; if (LIST_FIRST(&bp->b_dep) != NULL) buf_deallocate(bp); if (bp->b_flags & B_DELWRI) { --numdirtybuffers; numdirtywakeup(lodirtybuffers); } bp->b_flags &= ~(B_DELWRI | B_CACHE); if ((bp->b_flags & B_VMIO) == 0) { if (bp->b_bufsize) allocbuf(bp, 0); if (bp->b_vp) brelvp(bp); } } /* * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() * is called with B_DELWRI set, the underlying pages may wind up * getting freed causing a previous write (bdwrite()) to get 'lost' * because pages associated with a B_DELWRI bp are marked clean. * * We still allow the B_INVAL case to call vfs_vmio_release(), even * if B_DELWRI is set. * * If B_DELWRI is not set we may have to set B_RELBUF if we are low * on pages to return pages to the VM page queues. */ if (bp->b_flags & B_DELWRI) bp->b_flags &= ~B_RELBUF; else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) bp->b_flags |= B_RELBUF; /* * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer * constituted, not even NFS buffers now. Two flags effect this. If * B_INVAL, the struct buf is invalidated but the VM object is kept * around ( i.e. so it is trivial to reconstitute the buffer later ). * * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be * invalidated. BIO_ERROR cannot be set for a failed write unless the * buffer is also B_INVAL because it hits the re-dirtying code above. * * Normally we can do this whether a buffer is B_DELWRI or not. If * the buffer is an NFS buffer, it is tracking piecemeal writes or * the commit state and we cannot afford to lose the buffer. If the * buffer has a background write in progress, we need to keep it * around to prevent it from being reconstituted and starting a second * background write. */ if ((bp->b_flags & B_VMIO) && !(bp->b_vp->v_tag == VT_NFS && !vn_isdisk(bp->b_vp, NULL) && (bp->b_flags & B_DELWRI)) ) { int i, j, resid; vm_page_t m; off_t foff; vm_pindex_t poff; vm_object_t obj; struct vnode *vp; vp = bp->b_vp; obj = bp->b_object; /* * Get the base offset and length of the buffer. Note that * in the VMIO case if the buffer block size is not * page-aligned then b_data pointer may not be page-aligned. * But our b_pages[] array *IS* page aligned. * * block sizes less then DEV_BSIZE (usually 512) are not * supported due to the page granularity bits (m->valid, * m->dirty, etc...). * * See man buf(9) for more information */ resid = bp->b_bufsize; foff = bp->b_offset; for (i = 0; i < bp->b_npages; i++) { int had_bogus = 0; m = bp->b_pages[i]; vm_page_flag_clear(m, PG_ZERO); /* * If we hit a bogus page, fixup *all* the bogus pages * now. */ if (m == bogus_page) { poff = OFF_TO_IDX(bp->b_offset); had_bogus = 1; for (j = i; j < bp->b_npages; j++) { vm_page_t mtmp; mtmp = bp->b_pages[j]; if (mtmp == bogus_page) { mtmp = vm_page_lookup(obj, poff + j); if (!mtmp) { panic("brelse: page missing\n"); } bp->b_pages[j] = mtmp; } } if ((bp->b_flags & B_INVAL) == 0) { pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } m = bp->b_pages[i]; } if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { int poffset = foff & PAGE_MASK; int presid = resid > (PAGE_SIZE - poffset) ? (PAGE_SIZE - poffset) : resid; KASSERT(presid >= 0, ("brelse: extra page")); vm_page_set_invalid(m, poffset, presid); if (had_bogus) printf("avoided corruption bug in bogus_page/brelse code\n"); } resid -= PAGE_SIZE - (foff & PAGE_MASK); foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; } if (bp->b_flags & (B_INVAL | B_RELBUF)) vfs_vmio_release(bp); } else if (bp->b_flags & B_VMIO) { if (bp->b_flags & (B_INVAL | B_RELBUF)) { vfs_vmio_release(bp); } } if (bp->b_qindex != QUEUE_NONE) panic("brelse: free buffer onto another queue???"); if (BUF_REFCNT(bp) > 1) { /* do not release to free list */ BUF_UNLOCK(bp); splx(s); return; } /* enqueue */ /* buffers with no memory */ if (bp->b_bufsize == 0) { bp->b_flags |= B_INVAL; bp->b_xflags &= ~BX_BKGRDWRITE; if (bp->b_xflags & BX_BKGRDINPROG) panic("losing buffer 1"); if (bp->b_kvasize) { bp->b_qindex = QUEUE_EMPTYKVA; } else { bp->b_qindex = QUEUE_EMPTY; } TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); LIST_REMOVE(bp, b_hash); LIST_INSERT_HEAD(&invalhash, bp, b_hash); bp->b_dev = NODEV; /* buffers with junk contents */ } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { bp->b_flags |= B_INVAL; bp->b_xflags &= ~BX_BKGRDWRITE; if (bp->b_xflags & BX_BKGRDINPROG) panic("losing buffer 2"); bp->b_qindex = QUEUE_CLEAN; TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); LIST_REMOVE(bp, b_hash); LIST_INSERT_HEAD(&invalhash, bp, b_hash); bp->b_dev = NODEV; /* buffers that are locked */ } else if (bp->b_flags & B_LOCKED) { bp->b_qindex = QUEUE_LOCKED; TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); /* remaining buffers */ } else { if (bp->b_flags & B_DELWRI) bp->b_qindex = QUEUE_DIRTY; else bp->b_qindex = QUEUE_CLEAN; if (bp->b_flags & B_AGE) TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); else TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); } /* * If B_INVAL, clear B_DELWRI. We've already placed the buffer * on the correct queue. */ if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) bundirty(bp); /* * Fixup numfreebuffers count. The bp is on an appropriate queue * unless locked. We then bump numfreebuffers if it is not B_DELWRI. * We've already handled the B_INVAL case ( B_DELWRI will be clear * if B_INVAL is set ). */ if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) bufcountwakeup(); /* * Something we can maybe free or reuse */ if (bp->b_bufsize || bp->b_kvasize) bufspacewakeup(); /* unlock */ BUF_UNLOCK(bp); bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT | B_NOWDRAIN); if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("brelse: not dirty"); splx(s); } /* * Release a buffer back to the appropriate queue but do not try to free * it. The buffer is expected to be used again soon. * * bqrelse() is used by bdwrite() to requeue a delayed write, and used by * biodone() to requeue an async I/O on completion. It is also used when * known good buffers need to be requeued but we think we may need the data * again soon. * * XXX we should be able to leave the B_RELBUF hint set on completion. */ void bqrelse(struct buf * bp) { int s; s = splbio(); KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); if (bp->b_qindex != QUEUE_NONE) panic("bqrelse: free buffer onto another queue???"); if (BUF_REFCNT(bp) > 1) { /* do not release to free list */ BUF_UNLOCK(bp); splx(s); return; } if (bp->b_flags & B_LOCKED) { bp->b_ioflags &= ~BIO_ERROR; bp->b_qindex = QUEUE_LOCKED; TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); /* buffers with stale but valid contents */ } else if (bp->b_flags & B_DELWRI) { bp->b_qindex = QUEUE_DIRTY; TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); } else if (vm_page_count_severe()) { /* * We are too low on memory, we have to try to free the * buffer (most importantly: the wired pages making up its * backing store) *now*. */ splx(s); brelse(bp); return; } else { bp->b_qindex = QUEUE_CLEAN; TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); } if ((bp->b_flags & B_LOCKED) == 0 && ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { bufcountwakeup(); } /* * Something we can maybe free or reuse. */ if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) bufspacewakeup(); /* unlock */ BUF_UNLOCK(bp); bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("bqrelse: not dirty"); splx(s); } /* Give pages used by the bp back to the VM system (where possible) */ static void vfs_vmio_release(bp) struct buf *bp; { int i; vm_page_t m; GIANT_REQUIRED; for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; bp->b_pages[i] = NULL; /* * In order to keep page LRU ordering consistent, put * everything on the inactive queue. */ vm_page_unwire(m, 0); /* * We don't mess with busy pages, it is * the responsibility of the process that * busied the pages to deal with them. */ if ((m->flags & PG_BUSY) || (m->busy != 0)) continue; if (m->wire_count == 0) { vm_page_flag_clear(m, PG_ZERO); /* * Might as well free the page if we can and it has * no valid data. We also free the page if the * buffer was used for direct I/O */ if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { vm_page_busy(m); vm_page_protect(m, VM_PROT_NONE); vm_page_free(m); } else if (bp->b_flags & B_DIRECT) { vm_page_try_to_free(m); } else if (vm_page_count_severe()) { vm_page_try_to_cache(m); } } } pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); if (bp->b_bufsize) { bufspacewakeup(); bp->b_bufsize = 0; } bp->b_npages = 0; bp->b_flags &= ~B_VMIO; if (bp->b_vp) brelvp(bp); } /* * Check to see if a block is currently memory resident. */ struct buf * gbincore(struct vnode * vp, daddr_t blkno) { struct buf *bp; struct bufhashhdr *bh; bh = bufhash(vp, blkno); /* Search hash chain */ LIST_FOREACH(bp, bh, b_hash) { /* hit */ if (bp->b_vp == vp && bp->b_lblkno == blkno && (bp->b_flags & B_INVAL) == 0) { break; } } return (bp); } /* * vfs_bio_awrite: * * Implement clustered async writes for clearing out B_DELWRI buffers. * This is much better then the old way of writing only one buffer at * a time. Note that we may not be presented with the buffers in the * correct order, so we search for the cluster in both directions. */ int vfs_bio_awrite(struct buf * bp) { int i; int j; daddr_t lblkno = bp->b_lblkno; struct vnode *vp = bp->b_vp; int s; int ncl; struct buf *bpa; int nwritten; int size; int maxcl; s = splbio(); /* * right now we support clustered writing only to regular files. If * we find a clusterable block we could be in the middle of a cluster * rather then at the beginning. */ if ((vp->v_type == VREG) && (vp->v_mount != 0) && /* Only on nodes that have the size info */ (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { size = vp->v_mount->mnt_stat.f_iosize; maxcl = MAXPHYS / size; for (i = 1; i < maxcl; i++) { if ((bpa = gbincore(vp, lblkno + i)) && BUF_REFCNT(bpa) == 0 && ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == (B_DELWRI | B_CLUSTEROK)) && (bpa->b_bufsize == size)) { if ((bpa->b_blkno == bpa->b_lblkno) || (bpa->b_blkno != bp->b_blkno + ((i * size) >> DEV_BSHIFT))) break; } else { break; } } for (j = 1; i + j <= maxcl && j <= lblkno; j++) { if ((bpa = gbincore(vp, lblkno - j)) && BUF_REFCNT(bpa) == 0 && ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == (B_DELWRI | B_CLUSTEROK)) && (bpa->b_bufsize == size)) { if ((bpa->b_blkno == bpa->b_lblkno) || (bpa->b_blkno != bp->b_blkno - ((j * size) >> DEV_BSHIFT))) break; } else { break; } } --j; ncl = i + j; /* * this is a possible cluster write */ if (ncl != 1) { nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); splx(s); return nwritten; } } BUF_LOCK(bp, LK_EXCLUSIVE); bremfree(bp); bp->b_flags |= B_ASYNC; splx(s); /* * default (old) behavior, writing out only one block * * XXX returns b_bufsize instead of b_bcount for nwritten? */ nwritten = bp->b_bufsize; (void) BUF_WRITE(bp); return nwritten; } /* * getnewbuf: * * Find and initialize a new buffer header, freeing up existing buffers * in the bufqueues as necessary. The new buffer is returned locked. * * Important: B_INVAL is not set. If the caller wishes to throw the * buffer away, the caller must set B_INVAL prior to calling brelse(). * * We block if: * We have insufficient buffer headers * We have insufficient buffer space * buffer_map is too fragmented ( space reservation fails ) * If we have to flush dirty buffers ( but we try to avoid this ) * * To avoid VFS layer recursion we do not flush dirty buffers ourselves. * Instead we ask the buf daemon to do it for us. We attempt to * avoid piecemeal wakeups of the pageout daemon. */ static struct buf * getnewbuf(int slpflag, int slptimeo, int size, int maxsize) { struct buf *bp; struct buf *nbp; int defrag = 0; int nqindex; static int flushingbufs; GIANT_REQUIRED; /* * We can't afford to block since we might be holding a vnode lock, * which may prevent system daemons from running. We deal with * low-memory situations by proactively returning memory and running * async I/O rather then sync I/O. */ ++getnewbufcalls; --getnewbufrestarts; restart: ++getnewbufrestarts; /* * Setup for scan. If we do not have enough free buffers, * we setup a degenerate case that immediately fails. Note * that if we are specially marked process, we are allowed to * dip into our reserves. * * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN * * We start with EMPTYKVA. If the list is empty we backup to EMPTY. * However, there are a number of cases (defragging, reusing, ...) * where we cannot backup. */ nqindex = QUEUE_EMPTYKVA; nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); if (nbp == NULL) { /* * If no EMPTYKVA buffers and we are either * defragging or reusing, locate a CLEAN buffer * to free or reuse. If bufspace useage is low * skip this step so we can allocate a new buffer. */ if (defrag || bufspace >= lobufspace) { nqindex = QUEUE_CLEAN; nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); } /* * If we could not find or were not allowed to reuse a * CLEAN buffer, check to see if it is ok to use an EMPTY * buffer. We can only use an EMPTY buffer if allocating * its KVA would not otherwise run us out of buffer space. */ if (nbp == NULL && defrag == 0 && bufspace + maxsize < hibufspace) { nqindex = QUEUE_EMPTY; nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); } } /* * Run scan, possibly freeing data and/or kva mappings on the fly * depending. */ while ((bp = nbp) != NULL) { int qindex = nqindex; /* * Calculate next bp ( we can only use it if we do not block * or do other fancy things ). */ if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { switch(qindex) { case QUEUE_EMPTY: nqindex = QUEUE_EMPTYKVA; if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) break; /* fall through */ case QUEUE_EMPTYKVA: nqindex = QUEUE_CLEAN; if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) break; /* fall through */ case QUEUE_CLEAN: /* * nbp is NULL. */ break; } } /* * Sanity Checks */ KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); /* * Note: we no longer distinguish between VMIO and non-VMIO * buffers. */ KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); /* * If we are defragging then we need a buffer with * b_kvasize != 0. XXX this situation should no longer * occur, if defrag is non-zero the buffer's b_kvasize * should also be non-zero at this point. XXX */ if (defrag && bp->b_kvasize == 0) { printf("Warning: defrag empty buffer %p\n", bp); continue; } /* * Start freeing the bp. This is somewhat involved. nbp * remains valid only for QUEUE_EMPTY[KVA] bp's. */ if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) panic("getnewbuf: locked buf"); bremfree(bp); if (qindex == QUEUE_CLEAN) { if (bp->b_flags & B_VMIO) { bp->b_flags &= ~B_ASYNC; vfs_vmio_release(bp); } if (bp->b_vp) brelvp(bp); } /* * NOTE: nbp is now entirely invalid. We can only restart * the scan from this point on. * * Get the rest of the buffer freed up. b_kva* is still * valid after this operation. */ if (bp->b_rcred != NOCRED) { crfree(bp->b_rcred); bp->b_rcred = NOCRED; } if (bp->b_wcred != NOCRED) { crfree(bp->b_wcred); bp->b_wcred = NOCRED; } if (LIST_FIRST(&bp->b_dep) != NULL) buf_deallocate(bp); if (bp->b_xflags & BX_BKGRDINPROG) panic("losing buffer 3"); LIST_REMOVE(bp, b_hash); LIST_INSERT_HEAD(&invalhash, bp, b_hash); if (bp->b_bufsize) allocbuf(bp, 0); bp->b_flags = 0; bp->b_ioflags = 0; bp->b_xflags = 0; bp->b_dev = NODEV; bp->b_vp = NULL; bp->b_blkno = bp->b_lblkno = 0; bp->b_offset = NOOFFSET; bp->b_iodone = 0; bp->b_error = 0; bp->b_resid = 0; bp->b_bcount = 0; bp->b_npages = 0; bp->b_dirtyoff = bp->b_dirtyend = 0; bp->b_magic = B_MAGIC_BIO; bp->b_op = &buf_ops_bio; bp->b_object = NULL; LIST_INIT(&bp->b_dep); /* * If we are defragging then free the buffer. */ if (defrag) { bp->b_flags |= B_INVAL; bfreekva(bp); brelse(bp); defrag = 0; goto restart; } /* * If we are overcomitted then recover the buffer and its * KVM space. This occurs in rare situations when multiple * processes are blocked in getnewbuf() or allocbuf(). */ if (bufspace >= hibufspace) flushingbufs = 1; if (flushingbufs && bp->b_kvasize != 0) { bp->b_flags |= B_INVAL; bfreekva(bp); brelse(bp); goto restart; } if (bufspace < lobufspace) flushingbufs = 0; break; } /* * If we exhausted our list, sleep as appropriate. We may have to * wakeup various daemons and write out some dirty buffers. * * Generally we are sleeping due to insufficient buffer space. */ if (bp == NULL) { int flags; char *waitmsg; if (defrag) { flags = VFS_BIO_NEED_BUFSPACE; waitmsg = "nbufkv"; } else if (bufspace >= hibufspace) { waitmsg = "nbufbs"; flags = VFS_BIO_NEED_BUFSPACE; } else { waitmsg = "newbuf"; flags = VFS_BIO_NEED_ANY; } bd_speedup(); /* heeeelp */ needsbuffer |= flags; while (needsbuffer & flags) { if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) return (NULL); } } else { /* * We finally have a valid bp. We aren't quite out of the * woods, we still have to reserve kva space. In order * to keep fragmentation sane we only allocate kva in * BKVASIZE chunks. */ maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; if (maxsize != bp->b_kvasize) { vm_offset_t addr = 0; bfreekva(bp); if (vm_map_findspace(buffer_map, vm_map_min(buffer_map), maxsize, &addr)) { /* * Uh oh. Buffer map is to fragmented. We * must defragment the map. */ ++bufdefragcnt; defrag = 1; bp->b_flags |= B_INVAL; brelse(bp); goto restart; } if (addr) { vm_map_insert(buffer_map, NULL, 0, addr, addr + maxsize, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); bp->b_kvabase = (caddr_t) addr; bp->b_kvasize = maxsize; bufspace += bp->b_kvasize; ++bufreusecnt; } } bp->b_data = bp->b_kvabase; } return(bp); } /* * buf_daemon: * * buffer flushing daemon. Buffers are normally flushed by the * update daemon but if it cannot keep up this process starts to * take the load in an attempt to prevent getnewbuf() from blocking. */ static struct proc *bufdaemonproc; static struct kproc_desc buf_kp = { "bufdaemon", buf_daemon, &bufdaemonproc }; SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) static void buf_daemon() { int s; mtx_lock(&Giant); /* * This process needs to be suspended prior to shutdown sync. */ EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, SHUTDOWN_PRI_LAST); /* * This process is allowed to take the buffer cache to the limit */ s = splbio(); for (;;) { kthread_suspend_check(bufdaemonproc); bd_request = 0; /* * Do the flush. Limit the amount of in-transit I/O we * allow to build up, otherwise we would completely saturate * the I/O system. Wakeup any waiting processes before we * normally would so they can run in parallel with our drain. */ while (numdirtybuffers > lodirtybuffers) { if (flushbufqueues() == 0) break; waitrunningbufspace(); numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); } /* * Only clear bd_request if we have reached our low water * mark. The buf_daemon normally waits 1 second and * then incrementally flushes any dirty buffers that have * built up, within reason. * * If we were unable to hit our low water mark and couldn't * find any flushable buffers, we sleep half a second. * Otherwise we loop immediately. */ if (numdirtybuffers <= lodirtybuffers) { /* * We reached our low water mark, reset the * request and sleep until we are needed again. * The sleep is just so the suspend code works. */ bd_request = 0; tsleep(&bd_request, PVM, "psleep", hz); } else { /* * We couldn't find any flushable dirty buffers but * still have too many dirty buffers, we * have to sleep and try again. (rare) */ tsleep(&bd_request, PVM, "qsleep", hz / 2); } } } /* * flushbufqueues: * * Try to flush a buffer in the dirty queue. We must be careful to * free up B_INVAL buffers instead of write them, which NFS is * particularly sensitive to. */ static int flushbufqueues(void) { struct buf *bp; int r = 0; bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); while (bp) { KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); if ((bp->b_flags & B_DELWRI) != 0 && (bp->b_xflags & BX_BKGRDINPROG) == 0) { if (bp->b_flags & B_INVAL) { if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) panic("flushbufqueues: locked buf"); bremfree(bp); brelse(bp); ++r; break; } if (LIST_FIRST(&bp->b_dep) != NULL && (bp->b_flags & B_DEFERRED) == 0 && buf_countdeps(bp, 0)) { TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], bp, b_freelist); TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); bp->b_flags |= B_DEFERRED; bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); continue; } vfs_bio_awrite(bp); ++r; break; } bp = TAILQ_NEXT(bp, b_freelist); } return (r); } /* * Check to see if a block is currently memory resident. */ struct buf * incore(struct vnode * vp, daddr_t blkno) { struct buf *bp; int s = splbio(); bp = gbincore(vp, blkno); splx(s); return (bp); } /* * Returns true if no I/O is needed to access the * associated VM object. This is like incore except * it also hunts around in the VM system for the data. */ int inmem(struct vnode * vp, daddr_t blkno) { vm_object_t obj; vm_offset_t toff, tinc, size; vm_page_t m; vm_ooffset_t off; GIANT_REQUIRED; if (incore(vp, blkno)) return 1; if (vp->v_mount == NULL) return 0; if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) return 0; size = PAGE_SIZE; if (size > vp->v_mount->mnt_stat.f_iosize) size = vp->v_mount->mnt_stat.f_iosize; off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); if (!m) goto notinmem; tinc = size; if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); if (vm_page_is_valid(m, (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) goto notinmem; } return 1; notinmem: return (0); } /* * vfs_setdirty: * * Sets the dirty range for a buffer based on the status of the dirty * bits in the pages comprising the buffer. * * The range is limited to the size of the buffer. * * This routine is primarily used by NFS, but is generalized for the * B_VMIO case. */ static void vfs_setdirty(struct buf *bp) { int i; vm_object_t object; GIANT_REQUIRED; /* * Degenerate case - empty buffer */ if (bp->b_bufsize == 0) return; /* * We qualify the scan for modified pages on whether the * object has been flushed yet. The OBJ_WRITEABLE flag * is not cleared simply by protecting pages off. */ if ((bp->b_flags & B_VMIO) == 0) return; object = bp->b_pages[0]->object; if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) printf("Warning: object %p writeable but not mightbedirty\n", object); if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) printf("Warning: object %p mightbedirty but not writeable\n", object); if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { vm_offset_t boffset; vm_offset_t eoffset; /* * test the pages to see if they have been modified directly * by users through the VM system. */ for (i = 0; i < bp->b_npages; i++) { vm_page_flag_clear(bp->b_pages[i], PG_ZERO); vm_page_test_dirty(bp->b_pages[i]); } /* * Calculate the encompassing dirty range, boffset and eoffset, * (eoffset - boffset) bytes. */ for (i = 0; i < bp->b_npages; i++) { if (bp->b_pages[i]->dirty) break; } boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); for (i = bp->b_npages - 1; i >= 0; --i) { if (bp->b_pages[i]->dirty) { break; } } eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); /* * Fit it to the buffer. */ if (eoffset > bp->b_bcount) eoffset = bp->b_bcount; /* * If we have a good dirty range, merge with the existing * dirty range. */ if (boffset < eoffset) { if (bp->b_dirtyoff > boffset) bp->b_dirtyoff = boffset; if (bp->b_dirtyend < eoffset) bp->b_dirtyend = eoffset; } } } /* * getblk: * * Get a block given a specified block and offset into a file/device. * The buffers B_DONE bit will be cleared on return, making it almost * ready for an I/O initiation. B_INVAL may or may not be set on * return. The caller should clear B_INVAL prior to initiating a * READ. * * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for * an existing buffer. * * For a VMIO buffer, B_CACHE is modified according to the backing VM. * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set * and then cleared based on the backing VM. If the previous buffer is * non-0-sized but invalid, B_CACHE will be cleared. * * If getblk() must create a new buffer, the new buffer is returned with * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which * case it is returned with B_INVAL clear and B_CACHE set based on the * backing VM. * * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos * B_CACHE bit is clear. * * What this means, basically, is that the caller should use B_CACHE to * determine whether the buffer is fully valid or not and should clear * B_INVAL prior to issuing a read. If the caller intends to validate * the buffer by loading its data area with something, the caller needs * to clear B_INVAL. If the caller does this without issuing an I/O, * the caller should set B_CACHE ( as an optimization ), else the caller * should issue the I/O and biodone() will set B_CACHE if the I/O was * a write attempt or if it was a successfull read. If the caller * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR * prior to issuing the READ. biodone() will *not* clear B_INVAL. */ struct buf * getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) { struct buf *bp; int s; struct bufhashhdr *bh; if (size > MAXBSIZE) panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); s = splbio(); loop: /* * Block if we are low on buffers. Certain processes are allowed * to completely exhaust the buffer cache. * * If this check ever becomes a bottleneck it may be better to * move it into the else, when gbincore() fails. At the moment * it isn't a problem. * * XXX remove if 0 sections (clean this up after its proven) */ if (numfreebuffers == 0) { if (curthread == PCPU_GET(idlethread)) return NULL; needsbuffer |= VFS_BIO_NEED_ANY; } if ((bp = gbincore(vp, blkno))) { /* * Buffer is in-core. If the buffer is not busy, it must * be on a queue. */ if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, "getblk", slpflag, slptimeo) == ENOLCK) goto loop; splx(s); return (struct buf *) NULL; } /* * The buffer is locked. B_CACHE is cleared if the buffer is * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set * and for a VMIO buffer B_CACHE is adjusted according to the * backing VM cache. */ if (bp->b_flags & B_INVAL) bp->b_flags &= ~B_CACHE; else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) bp->b_flags |= B_CACHE; bremfree(bp); /* * check for size inconsistancies for non-VMIO case. */ if (bp->b_bcount != size) { if ((bp->b_flags & B_VMIO) == 0 || (size > bp->b_kvasize)) { if (bp->b_flags & B_DELWRI) { bp->b_flags |= B_NOCACHE; BUF_WRITE(bp); } else { if ((bp->b_flags & B_VMIO) && (LIST_FIRST(&bp->b_dep) == NULL)) { bp->b_flags |= B_RELBUF; brelse(bp); } else { bp->b_flags |= B_NOCACHE; BUF_WRITE(bp); } } goto loop; } } /* * If the size is inconsistant in the VMIO case, we can resize * the buffer. This might lead to B_CACHE getting set or * cleared. If the size has not changed, B_CACHE remains * unchanged from its previous state. */ if (bp->b_bcount != size) allocbuf(bp, size); KASSERT(bp->b_offset != NOOFFSET, ("getblk: no buffer offset")); /* * A buffer with B_DELWRI set and B_CACHE clear must * be committed before we can return the buffer in * order to prevent the caller from issuing a read * ( due to B_CACHE not being set ) and overwriting * it. * * Most callers, including NFS and FFS, need this to * operate properly either because they assume they * can issue a read if B_CACHE is not set, or because * ( for example ) an uncached B_DELWRI might loop due * to softupdates re-dirtying the buffer. In the latter * case, B_CACHE is set after the first write completes, * preventing further loops. * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE * above while extending the buffer, we cannot allow the * buffer to remain with B_CACHE set after the write * completes or it will represent a corrupt state. To * deal with this we set B_NOCACHE to scrap the buffer * after the write. * * We might be able to do something fancy, like setting * B_CACHE in bwrite() except if B_DELWRI is already set, * so the below call doesn't set B_CACHE, but that gets real * confusing. This is much easier. */ if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { bp->b_flags |= B_NOCACHE; BUF_WRITE(bp); goto loop; } splx(s); bp->b_flags &= ~B_DONE; } else { /* * Buffer is not in-core, create new buffer. The buffer * returned by getnewbuf() is locked. Note that the returned * buffer is also considered valid (not marked B_INVAL). */ int bsize, maxsize, vmio; off_t offset; if (vn_isdisk(vp, NULL)) bsize = DEV_BSIZE; else if (vp->v_mountedhere) bsize = vp->v_mountedhere->mnt_stat.f_iosize; else if (vp->v_mount) bsize = vp->v_mount->mnt_stat.f_iosize; else bsize = size; offset = blkno * bsize; vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); maxsize = vmio ? size + (offset & PAGE_MASK) : size; maxsize = imax(maxsize, bsize); if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { if (slpflag || slptimeo) { splx(s); return NULL; } goto loop; } /* * This code is used to make sure that a buffer is not * created while the getnewbuf routine is blocked. * This can be a problem whether the vnode is locked or not. * If the buffer is created out from under us, we have to * throw away the one we just created. There is now window * race because we are safely running at splbio() from the * point of the duplicate buffer creation through to here, * and we've locked the buffer. */ if (gbincore(vp, blkno)) { bp->b_flags |= B_INVAL; brelse(bp); goto loop; } /* * Insert the buffer into the hash, so that it can * be found by incore. */ bp->b_blkno = bp->b_lblkno = blkno; bp->b_offset = offset; bgetvp(vp, bp); LIST_REMOVE(bp, b_hash); bh = bufhash(vp, blkno); LIST_INSERT_HEAD(bh, bp, b_hash); /* * set B_VMIO bit. allocbuf() the buffer bigger. Since the * buffer size starts out as 0, B_CACHE will be set by * allocbuf() for the VMIO case prior to it testing the * backing store for validity. */ if (vmio) { bp->b_flags |= B_VMIO; #if defined(VFS_BIO_DEBUG) if (vp->v_type != VREG) printf("getblk: vmioing file type %d???\n", vp->v_type); #endif VOP_GETVOBJECT(vp, &bp->b_object); - vm_object_reference(bp->b_object); } else { bp->b_flags &= ~B_VMIO; bp->b_object = NULL; } allocbuf(bp, size); splx(s); bp->b_flags &= ~B_DONE; } return (bp); } /* * Get an empty, disassociated buffer of given size. The buffer is initially * set to B_INVAL. */ struct buf * geteblk(int size) { struct buf *bp; int s; int maxsize; maxsize = (size + BKVAMASK) & ~BKVAMASK; s = splbio(); while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); splx(s); allocbuf(bp, size); bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ return (bp); } /* * This code constitutes the buffer memory from either anonymous system * memory (in the case of non-VMIO operations) or from an associated * VM object (in the case of VMIO operations). This code is able to * resize a buffer up or down. * * Note that this code is tricky, and has many complications to resolve * deadlock or inconsistant data situations. Tread lightly!!! * There are B_CACHE and B_DELWRI interactions that must be dealt with by * the caller. Calling this code willy nilly can result in the loss of data. * * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with * B_CACHE for the non-VMIO case. */ int allocbuf(struct buf *bp, int size) { int newbsize, mbsize; int i; GIANT_REQUIRED; if (BUF_REFCNT(bp) == 0) panic("allocbuf: buffer not busy"); if (bp->b_kvasize < size) panic("allocbuf: buffer too small"); if ((bp->b_flags & B_VMIO) == 0) { caddr_t origbuf; int origbufsize; /* * Just get anonymous memory from the kernel. Don't * mess with B_CACHE. */ mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); if (bp->b_flags & B_MALLOC) newbsize = mbsize; else newbsize = round_page(size); if (newbsize < bp->b_bufsize) { /* * malloced buffers are not shrunk */ if (bp->b_flags & B_MALLOC) { if (newbsize) { bp->b_bcount = size; } else { free(bp->b_data, M_BIOBUF); if (bp->b_bufsize) { bufmallocspace -= bp->b_bufsize; bufspacewakeup(); bp->b_bufsize = 0; } bp->b_data = bp->b_kvabase; bp->b_bcount = 0; bp->b_flags &= ~B_MALLOC; } return 1; } vm_hold_free_pages( bp, (vm_offset_t) bp->b_data + newbsize, (vm_offset_t) bp->b_data + bp->b_bufsize); } else if (newbsize > bp->b_bufsize) { /* * We only use malloced memory on the first allocation. * and revert to page-allocated memory when the buffer * grows. */ if ( (bufmallocspace < maxbufmallocspace) && (bp->b_bufsize == 0) && (mbsize <= PAGE_SIZE/2)) { bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); bp->b_bufsize = mbsize; bp->b_bcount = size; bp->b_flags |= B_MALLOC; bufmallocspace += mbsize; return 1; } origbuf = NULL; origbufsize = 0; /* * If the buffer is growing on its other-than-first allocation, * then we revert to the page-allocation scheme. */ if (bp->b_flags & B_MALLOC) { origbuf = bp->b_data; origbufsize = bp->b_bufsize; bp->b_data = bp->b_kvabase; if (bp->b_bufsize) { bufmallocspace -= bp->b_bufsize; bufspacewakeup(); bp->b_bufsize = 0; } bp->b_flags &= ~B_MALLOC; newbsize = round_page(newbsize); } vm_hold_load_pages( bp, (vm_offset_t) bp->b_data + bp->b_bufsize, (vm_offset_t) bp->b_data + newbsize); if (origbuf) { bcopy(origbuf, bp->b_data, origbufsize); free(origbuf, M_BIOBUF); } } } else { vm_page_t m; int desiredpages; newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); desiredpages = (size == 0) ? 0 : num_pages((bp->b_offset & PAGE_MASK) + newbsize); if (bp->b_flags & B_MALLOC) panic("allocbuf: VMIO buffer can't be malloced"); /* * Set B_CACHE initially if buffer is 0 length or will become * 0-length. */ if (size == 0 || bp->b_bufsize == 0) bp->b_flags |= B_CACHE; if (newbsize < bp->b_bufsize) { /* * DEV_BSIZE aligned new buffer size is less then the * DEV_BSIZE aligned existing buffer size. Figure out * if we have to remove any pages. */ if (desiredpages < bp->b_npages) { for (i = desiredpages; i < bp->b_npages; i++) { /* * the page is not freed here -- it * is the responsibility of * vnode_pager_setsize */ m = bp->b_pages[i]; KASSERT(m != bogus_page, ("allocbuf: bogus page found")); while (vm_page_sleep_busy(m, TRUE, "biodep")) ; bp->b_pages[i] = NULL; vm_page_unwire(m, 0); } pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); bp->b_npages = desiredpages; } } else if (size > bp->b_bcount) { /* * We are growing the buffer, possibly in a * byte-granular fashion. */ struct vnode *vp; vm_object_t obj; vm_offset_t toff; vm_offset_t tinc; /* * Step 1, bring in the VM pages from the object, * allocating them if necessary. We must clear * B_CACHE if these pages are not valid for the * range covered by the buffer. */ vp = bp->b_vp; obj = bp->b_object; while (bp->b_npages < desiredpages) { vm_page_t m; vm_pindex_t pi; pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; if ((m = vm_page_lookup(obj, pi)) == NULL) { /* * note: must allocate system pages * since blocking here could intefere * with paging I/O, no matter which * process we are. */ m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM); if (m == NULL) { VM_WAIT; vm_pageout_deficit += desiredpages - bp->b_npages; } else { vm_page_wire(m); vm_page_wakeup(m); bp->b_flags &= ~B_CACHE; bp->b_pages[bp->b_npages] = m; ++bp->b_npages; } continue; } /* * We found a page. If we have to sleep on it, * retry because it might have gotten freed out * from under us. * * We can only test PG_BUSY here. Blocking on * m->busy might lead to a deadlock: * * vm_fault->getpages->cluster_read->allocbuf * */ if (vm_page_sleep_busy(m, FALSE, "pgtblk")) continue; /* * We have a good page. Should we wakeup the * page daemon? */ if ((curproc != pageproc) && ((m->queue - m->pc) == PQ_CACHE) && ((cnt.v_free_count + cnt.v_cache_count) < (cnt.v_free_min + cnt.v_cache_min))) { pagedaemon_wakeup(); } vm_page_flag_clear(m, PG_ZERO); vm_page_wire(m); bp->b_pages[bp->b_npages] = m; ++bp->b_npages; } /* * Step 2. We've loaded the pages into the buffer, * we have to figure out if we can still have B_CACHE * set. Note that B_CACHE is set according to the * byte-granular range ( bcount and size ), new the * aligned range ( newbsize ). * * The VM test is against m->valid, which is DEV_BSIZE * aligned. Needless to say, the validity of the data * needs to also be DEV_BSIZE aligned. Note that this * fails with NFS if the server or some other client * extends the file's EOF. If our buffer is resized, * B_CACHE may remain set! XXX */ toff = bp->b_bcount; tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); while ((bp->b_flags & B_CACHE) && toff < size) { vm_pindex_t pi; if (tinc > (size - toff)) tinc = size - toff; pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; vfs_buf_test_cache( bp, bp->b_offset, toff, tinc, bp->b_pages[pi] ); toff += tinc; tinc = PAGE_SIZE; } /* * Step 3, fixup the KVM pmap. Remember that * bp->b_data is relative to bp->b_offset, but * bp->b_offset may be offset into the first page. */ bp->b_data = (caddr_t) trunc_page((vm_offset_t)bp->b_data); pmap_qenter( (vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages ); bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | (vm_offset_t)(bp->b_offset & PAGE_MASK)); } } if (newbsize < bp->b_bufsize) bufspacewakeup(); bp->b_bufsize = newbsize; /* actual buffer allocation */ bp->b_bcount = size; /* requested buffer size */ return 1; } /* * bufwait: * * Wait for buffer I/O completion, returning error status. The buffer * is left locked and B_DONE on return. B_EINTR is converted into a EINTR * error and cleared. */ int bufwait(register struct buf * bp) { int s; s = splbio(); while ((bp->b_flags & B_DONE) == 0) { if (bp->b_iocmd == BIO_READ) tsleep(bp, PRIBIO, "biord", 0); else tsleep(bp, PRIBIO, "biowr", 0); } splx(s); if (bp->b_flags & B_EINTR) { bp->b_flags &= ~B_EINTR; return (EINTR); } if (bp->b_ioflags & BIO_ERROR) { return (bp->b_error ? bp->b_error : EIO); } else { return (0); } } /* * Call back function from struct bio back up to struct buf. * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). */ void bufdonebio(struct bio *bp) { bufdone(bp->bio_caller2); } /* * bufdone: * * Finish I/O on a buffer, optionally calling a completion function. * This is usually called from an interrupt so process blocking is * not allowed. * * biodone is also responsible for setting B_CACHE in a B_VMIO bp. * In a non-VMIO bp, B_CACHE will be set on the next getblk() * assuming B_INVAL is clear. * * For the VMIO case, we set B_CACHE if the op was a read and no * read error occured, or if the op was a write. B_CACHE is never * set if the buffer is invalid or otherwise uncacheable. * * biodone does not mess with B_INVAL, allowing the I/O routine or the * initiator to leave B_INVAL set to brelse the buffer out of existance * in the biodone routine. */ void bufdone(struct buf *bp) { int s; void (*biodone)(struct buf *); GIANT_REQUIRED; s = splbio(); KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); bp->b_flags |= B_DONE; runningbufwakeup(bp); if (bp->b_iocmd == BIO_DELETE) { brelse(bp); splx(s); return; } if (bp->b_iocmd == BIO_WRITE) { vwakeup(bp); } /* call optional completion function if requested */ if (bp->b_iodone != NULL) { biodone = bp->b_iodone; bp->b_iodone = NULL; (*biodone) (bp); splx(s); return; } if (LIST_FIRST(&bp->b_dep) != NULL) buf_complete(bp); if (bp->b_flags & B_VMIO) { int i; vm_ooffset_t foff; vm_page_t m; vm_object_t obj; int iosize; struct vnode *vp = bp->b_vp; obj = bp->b_object; #if defined(VFS_BIO_DEBUG) if (vp->v_usecount == 0) { panic("biodone: zero vnode ref count"); } if ((vp->v_flag & VOBJBUF) == 0) { panic("biodone: vnode is not setup for merged cache"); } #endif foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("biodone: no buffer offset")); #if defined(VFS_BIO_DEBUG) if (obj->paging_in_progress < bp->b_npages) { printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", obj->paging_in_progress, bp->b_npages); } #endif /* * Set B_CACHE if the op was a normal read and no error * occured. B_CACHE is set for writes in the b*write() * routines. */ iosize = bp->b_bcount - bp->b_resid; if (bp->b_iocmd == BIO_READ && !(bp->b_flags & (B_INVAL|B_NOCACHE)) && !(bp->b_ioflags & BIO_ERROR)) { bp->b_flags |= B_CACHE; } for (i = 0; i < bp->b_npages; i++) { int bogusflag = 0; int resid; resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; if (resid > iosize) resid = iosize; /* * cleanup bogus pages, restoring the originals */ m = bp->b_pages[i]; if (m == bogus_page) { bogusflag = 1; m = vm_page_lookup(obj, OFF_TO_IDX(foff)); if (m == NULL) panic("biodone: page disappeared!"); bp->b_pages[i] = m; pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } #if defined(VFS_BIO_DEBUG) if (OFF_TO_IDX(foff) != m->pindex) { printf( "biodone: foff(%lu)/m->pindex(%d) mismatch\n", (unsigned long)foff, m->pindex); } #endif /* * In the write case, the valid and clean bits are * already changed correctly ( see bdwrite() ), so we * only need to do this here in the read case. */ if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { vfs_page_set_valid(bp, foff, i, m); } vm_page_flag_clear(m, PG_ZERO); /* * when debugging new filesystems or buffer I/O methods, this * is the most common error that pops up. if you see this, you * have not set the page busy flag correctly!!! */ if (m->busy == 0) { printf("biodone: page busy < 0, " "pindex: %d, foff: 0x(%x,%x), " "resid: %d, index: %d\n", (int) m->pindex, (int)(foff >> 32), (int) foff & 0xffffffff, resid, i); if (!vn_isdisk(vp, NULL)) printf(" iosize: %ld, lblkno: %jd, flags: 0x%lx, npages: %d\n", bp->b_vp->v_mount->mnt_stat.f_iosize, (intmax_t) bp->b_lblkno, bp->b_flags, bp->b_npages); else printf(" VDEV, lblkno: %jd, flags: 0x%lx, npages: %d\n", (intmax_t) bp->b_lblkno, bp->b_flags, bp->b_npages); printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", m->valid, m->dirty, m->wire_count); panic("biodone: page busy < 0\n"); } vm_page_io_finish(m); vm_object_pip_subtract(obj, 1); foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; iosize -= resid; } if (obj) vm_object_pip_wakeupn(obj, 0); } /* * For asynchronous completions, release the buffer now. The brelse * will do a wakeup there if necessary - so no need to do a wakeup * here in the async case. The sync case always needs to do a wakeup. */ if (bp->b_flags & B_ASYNC) { if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) brelse(bp); else bqrelse(bp); } else { wakeup(bp); } splx(s); } /* * This routine is called in lieu of iodone in the case of * incomplete I/O. This keeps the busy status for pages * consistant. */ void vfs_unbusy_pages(struct buf * bp) { int i; GIANT_REQUIRED; runningbufwakeup(bp); if (bp->b_flags & B_VMIO) { vm_object_t obj; obj = bp->b_object; for (i = 0; i < bp->b_npages; i++) { vm_page_t m = bp->b_pages[i]; if (m == bogus_page) { m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); if (!m) { panic("vfs_unbusy_pages: page missing\n"); } bp->b_pages[i] = m; pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } vm_object_pip_subtract(obj, 1); vm_page_flag_clear(m, PG_ZERO); vm_page_io_finish(m); } vm_object_pip_wakeupn(obj, 0); } } /* * vfs_page_set_valid: * * Set the valid bits in a page based on the supplied offset. The * range is restricted to the buffer's size. * * This routine is typically called after a read completes. */ static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) { vm_ooffset_t soff, eoff; GIANT_REQUIRED; /* * Start and end offsets in buffer. eoff - soff may not cross a * page boundry or cross the end of the buffer. The end of the * buffer, in this case, is our file EOF, not the allocation size * of the buffer. */ soff = off; eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; if (eoff > bp->b_offset + bp->b_bcount) eoff = bp->b_offset + bp->b_bcount; /* * Set valid range. This is typically the entire buffer and thus the * entire page. */ if (eoff > soff) { vm_page_set_validclean( m, (vm_offset_t) (soff & PAGE_MASK), (vm_offset_t) (eoff - soff) ); } } /* * This routine is called before a device strategy routine. * It is used to tell the VM system that paging I/O is in * progress, and treat the pages associated with the buffer * almost as being PG_BUSY. Also the object paging_in_progress * flag is handled to make sure that the object doesn't become * inconsistant. * * Since I/O has not been initiated yet, certain buffer flags * such as BIO_ERROR or B_INVAL may be in an inconsistant state * and should be ignored. */ void vfs_busy_pages(struct buf * bp, int clear_modify) { int i, bogus; GIANT_REQUIRED; if (bp->b_flags & B_VMIO) { vm_object_t obj; vm_ooffset_t foff; obj = bp->b_object; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_busy_pages: no buffer offset")); vfs_setdirty(bp); retry: for (i = 0; i < bp->b_npages; i++) { vm_page_t m = bp->b_pages[i]; if (vm_page_sleep_busy(m, FALSE, "vbpage")) goto retry; } bogus = 0; for (i = 0; i < bp->b_npages; i++) { vm_page_t m = bp->b_pages[i]; vm_page_flag_clear(m, PG_ZERO); if ((bp->b_flags & B_CLUSTER) == 0) { vm_object_pip_add(obj, 1); vm_page_io_start(m); } /* * When readying a buffer for a read ( i.e * clear_modify == 0 ), it is important to do * bogus_page replacement for valid pages in * partially instantiated buffers. Partially * instantiated buffers can, in turn, occur when * reconstituting a buffer from its VM backing store * base. We only have to do this if B_CACHE is * clear ( which causes the I/O to occur in the * first place ). The replacement prevents the read * I/O from overwriting potentially dirty VM-backed * pages. XXX bogus page replacement is, uh, bogus. * It may not work properly with small-block devices. * We need to find a better way. */ vm_page_protect(m, VM_PROT_NONE); if (clear_modify) vfs_page_set_valid(bp, foff, i, m); else if (m->valid == VM_PAGE_BITS_ALL && (bp->b_flags & B_CACHE) == 0) { bp->b_pages[i] = bogus_page; bogus++; } foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; } if (bogus) pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } } /* * Tell the VM system that the pages associated with this buffer * are clean. This is used for delayed writes where the data is * going to go to disk eventually without additional VM intevention. * * Note that while we only really need to clean through to b_bcount, we * just go ahead and clean through to b_bufsize. */ static void vfs_clean_pages(struct buf * bp) { int i; GIANT_REQUIRED; if (bp->b_flags & B_VMIO) { vm_ooffset_t foff; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_clean_pages: no buffer offset")); for (i = 0; i < bp->b_npages; i++) { vm_page_t m = bp->b_pages[i]; vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; vm_ooffset_t eoff = noff; if (eoff > bp->b_offset + bp->b_bufsize) eoff = bp->b_offset + bp->b_bufsize; vfs_page_set_valid(bp, foff, i, m); /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ foff = noff; } } } /* * vfs_bio_set_validclean: * * Set the range within the buffer to valid and clean. The range is * relative to the beginning of the buffer, b_offset. Note that b_offset * itself may be offset from the beginning of the first page. * */ void vfs_bio_set_validclean(struct buf *bp, int base, int size) { if (bp->b_flags & B_VMIO) { int i; int n; /* * Fixup base to be relative to beginning of first page. * Set initial n to be the maximum number of bytes in the * first page that can be validated. */ base += (bp->b_offset & PAGE_MASK); n = PAGE_SIZE - (base & PAGE_MASK); for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { vm_page_t m = bp->b_pages[i]; if (n > size) n = size; vm_page_set_validclean(m, base & PAGE_MASK, n); base += n; size -= n; n = PAGE_SIZE; } } } /* * vfs_bio_clrbuf: * * clear a buffer. This routine essentially fakes an I/O, so we need * to clear BIO_ERROR and B_INVAL. * * Note that while we only theoretically need to clear through b_bcount, * we go ahead and clear through b_bufsize. */ void vfs_bio_clrbuf(struct buf *bp) { int i, mask = 0; caddr_t sa, ea; GIANT_REQUIRED; if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && (bp->b_offset & PAGE_MASK) == 0) { mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; if ((bp->b_pages[0]->valid & mask) == mask) { bp->b_resid = 0; return; } if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && ((bp->b_pages[0]->valid & mask) == 0)) { bzero(bp->b_data, bp->b_bufsize); bp->b_pages[0]->valid |= mask; bp->b_resid = 0; return; } } ea = sa = bp->b_data; for(i=0;ib_npages;i++,sa=ea) { int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); ea = (caddr_t)(vm_offset_t)ulmin( (u_long)(vm_offset_t)ea, (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; if ((bp->b_pages[i]->valid & mask) == mask) continue; if ((bp->b_pages[i]->valid & mask) == 0) { if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { bzero(sa, ea - sa); } } else { for (; sa < ea; sa += DEV_BSIZE, j++) { if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && (bp->b_pages[i]->valid & (1<b_pages[i]->valid |= mask; vm_page_flag_clear(bp->b_pages[i], PG_ZERO); } bp->b_resid = 0; } else { clrbuf(bp); } } /* * vm_hold_load_pages and vm_hold_free_pages get pages into * a buffers address space. The pages are anonymous and are * not associated with a file object. */ static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) { vm_offset_t pg; vm_page_t p; int index; GIANT_REQUIRED; to = round_page(to); from = round_page(from); index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; for (pg = from; pg < to; pg += PAGE_SIZE, index++) { tryagain: /* * note: must allocate system pages since blocking here * could intefere with paging I/O, no matter which * process we are. */ p = vm_page_alloc(kernel_object, ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), VM_ALLOC_SYSTEM); if (!p) { vm_pageout_deficit += (to - from) >> PAGE_SHIFT; VM_WAIT; goto tryagain; } vm_page_wire(p); p->valid = VM_PAGE_BITS_ALL; vm_page_flag_clear(p, PG_ZERO); pmap_qenter(pg, &p, 1); bp->b_pages[index] = p; vm_page_wakeup(p); } bp->b_npages = index; } /* Return pages associated with this buf to the vm system */ void vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) { vm_offset_t pg; vm_page_t p; int index, newnpages; GIANT_REQUIRED; from = round_page(from); to = round_page(to); newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; for (pg = from; pg < to; pg += PAGE_SIZE, index++) { p = bp->b_pages[index]; if (p && (index < bp->b_npages)) { if (p->busy) { printf( "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); } bp->b_pages[index] = NULL; pmap_qremove(pg, 1); vm_page_busy(p); vm_page_unwire(p, 0); vm_page_free(p); } } bp->b_npages = newnpages; } #include "opt_ddb.h" #ifdef DDB #include /* DDB command to show buffer data */ DB_SHOW_COMMAND(buffer, db_show_buffer) { /* get args */ struct buf *bp = (struct buf *)addr; if (!have_addr) { db_printf("usage: show buffer \n"); return; } db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); db_printf( "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, major(bp->b_dev), minor(bp->b_dev), bp->b_data, (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); if (bp->b_npages) { int i; db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); for (i = 0; i < bp->b_npages; i++) { vm_page_t m; m = bp->b_pages[i]; db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); if ((i + 1) < bp->b_npages) db_printf(","); } db_printf("\n"); } } #endif /* DDB */ diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c index d8e960a2ba10..83c7bec307db 100644 --- a/sys/kern/vfs_subr.c +++ b/sys/kern/vfs_subr.c @@ -1,3135 +1,3133 @@ /* * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 * $FreeBSD$ */ /* * External virtual filesystem routines */ #include "opt_ddb.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); static void addalias(struct vnode *vp, dev_t nvp_rdev); static void insmntque(struct vnode *vp, struct mount *mp); static void vclean(struct vnode *vp, int flags, struct thread *td); static void vlruvp(struct vnode *vp); /* * Number of vnodes in existence. Increased whenever getnewvnode() * allocates a new vnode, never decreased. */ static unsigned long numvnodes; SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); /* * Conversion tables for conversion from vnode types to inode formats * and back. */ enum vtype iftovt_tab[16] = { VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, }; int vttoif_tab[9] = { 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, S_IFSOCK, S_IFIFO, S_IFMT, }; /* * List of vnodes that are ready for recycling. */ static TAILQ_HEAD(freelst, vnode) vnode_free_list; /* * Minimum number of free vnodes. If there are fewer than this free vnodes, * getnewvnode() will return a newly allocated vnode. */ static u_long wantfreevnodes = 25; SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); /* Number of vnodes in the free list. */ static u_long freevnodes; SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); /* * Various variables used for debugging the new implementation of * reassignbuf(). * XXX these are probably of (very) limited utility now. */ static int reassignbufcalls; SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); static int reassignbufloops; SYSCTL_INT(_vfs, OID_AUTO, reassignbufloops, CTLFLAG_RW, &reassignbufloops, 0, ""); static int reassignbufsortgood; SYSCTL_INT(_vfs, OID_AUTO, reassignbufsortgood, CTLFLAG_RW, &reassignbufsortgood, 0, ""); static int reassignbufsortbad; SYSCTL_INT(_vfs, OID_AUTO, reassignbufsortbad, CTLFLAG_RW, &reassignbufsortbad, 0, ""); /* Set to 0 for old insertion-sort based reassignbuf, 1 for modern method. */ static int reassignbufmethod = 1; SYSCTL_INT(_vfs, OID_AUTO, reassignbufmethod, CTLFLAG_RW, &reassignbufmethod, 0, ""); static int nameileafonly; SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, ""); #ifdef ENABLE_VFS_IOOPT /* See NOTES for a description of this setting. */ int vfs_ioopt; SYSCTL_INT(_vfs, OID_AUTO, ioopt, CTLFLAG_RW, &vfs_ioopt, 0, ""); #endif /* * Cache for the mount type id assigned to NFS. This is used for * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. */ int nfs_mount_type = -1; /* To keep more than one thread at a time from running vfs_getnewfsid */ static struct mtx mntid_mtx; /* For any iteration/modification of vnode_free_list */ static struct mtx vnode_free_list_mtx; /* * For any iteration/modification of dev->si_hlist (linked through * v_specnext) */ static struct mtx spechash_mtx; /* Publicly exported FS */ struct nfs_public nfs_pub; /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ static uma_zone_t vnode_zone; static uma_zone_t vnodepoll_zone; /* Set to 1 to print out reclaim of active vnodes */ int prtactive; /* * The workitem queue. * * It is useful to delay writes of file data and filesystem metadata * for tens of seconds so that quickly created and deleted files need * not waste disk bandwidth being created and removed. To realize this, * we append vnodes to a "workitem" queue. When running with a soft * updates implementation, most pending metadata dependencies should * not wait for more than a few seconds. Thus, mounted on block devices * are delayed only about a half the time that file data is delayed. * Similarly, directory updates are more critical, so are only delayed * about a third the time that file data is delayed. Thus, there are * SYNCER_MAXDELAY queues that are processed round-robin at a rate of * one each second (driven off the filesystem syncer process). The * syncer_delayno variable indicates the next queue that is to be processed. * Items that need to be processed soon are placed in this queue: * * syncer_workitem_pending[syncer_delayno] * * A delay of fifteen seconds is done by placing the request fifteen * entries later in the queue: * * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] * */ static int syncer_delayno; static long syncer_mask; LIST_HEAD(synclist, vnode); static struct synclist *syncer_workitem_pending; #define SYNCER_MAXDELAY 32 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ static int syncdelay = 30; /* max time to delay syncing data */ static int filedelay = 30; /* time to delay syncing files */ SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); static int dirdelay = 29; /* time to delay syncing directories */ SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); static int metadelay = 28; /* time to delay syncing metadata */ SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); static int rushjob; /* number of slots to run ASAP */ static int stat_rush_requests; /* number of times I/O speeded up */ SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); /* * Number of vnodes we want to exist at any one time. This is mostly used * to size hash tables in vnode-related code. It is normally not used in * getnewvnode(), as wantfreevnodes is normally nonzero.) * * XXX desiredvnodes is historical cruft and should not exist. */ int desiredvnodes; SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, &desiredvnodes, 0, "Maximum number of vnodes"); static int minvnodes; SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, &minvnodes, 0, "Minimum number of vnodes"); static int vnlru_nowhere; SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); /* Hook for calling soft updates */ int (*softdep_process_worklist_hook)(struct mount *); #ifdef DEBUG_VFS_LOCKS /* Print lock violations */ int vfs_badlock_print = 1; /* Panic on violation */ int vfs_badlock_panic = 1; void vop_rename_pre(void *ap) { struct vop_rename_args *a = ap; /* Check the source (from) */ if (a->a_tdvp != a->a_fdvp) ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked.\n"); if (a->a_tvp != a->a_fvp) ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked.\n"); /* Check the target */ if (a->a_tvp) ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked.\n"); ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked.\n"); } void vop_strategy_pre(void *ap) { struct vop_strategy_args *a = ap; int status; status = lockstatus(&a->a_bp->b_lock, curthread); if (status != LK_SHARED && status != LK_EXCLUSIVE) { if (vfs_badlock_print) printf("VOP_STRATEGY: bp is not locked but should be.\n"); if (vfs_badlock_panic) Debugger("Lock violation.\n"); } } #endif /* DEBUG_VFS_LOCKS */ void v_addpollinfo(struct vnode *vp) { vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK); mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); } /* * Initialize the vnode management data structures. */ static void vntblinit(void *dummy __unused) { desiredvnodes = maxproc + cnt.v_page_count / 4; minvnodes = desiredvnodes / 4; mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); mtx_init(&mntvnode_mtx, "mntvnode", NULL, MTX_DEF); mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF); TAILQ_INIT(&vnode_free_list); mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); /* * Initialize the filesystem syncer. */ syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, &syncer_mask); syncer_maxdelay = syncer_mask + 1; } SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) /* * Mark a mount point as busy. Used to synchronize access and to delay * unmounting. Interlock is not released on failure. */ int vfs_busy(mp, flags, interlkp, td) struct mount *mp; int flags; struct mtx *interlkp; struct thread *td; { int lkflags; if (mp->mnt_kern_flag & MNTK_UNMOUNT) { if (flags & LK_NOWAIT) return (ENOENT); mp->mnt_kern_flag |= MNTK_MWAIT; /* * Since all busy locks are shared except the exclusive * lock granted when unmounting, the only place that a * wakeup needs to be done is at the release of the * exclusive lock at the end of dounmount. */ msleep(mp, interlkp, PVFS, "vfs_busy", 0); return (ENOENT); } lkflags = LK_SHARED | LK_NOPAUSE; if (interlkp) lkflags |= LK_INTERLOCK; if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td)) panic("vfs_busy: unexpected lock failure"); return (0); } /* * Free a busy filesystem. */ void vfs_unbusy(mp, td) struct mount *mp; struct thread *td; { lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); } /* * Lookup a mount point by filesystem identifier. */ struct mount * vfs_getvfs(fsid) fsid_t *fsid; { register struct mount *mp; mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { mtx_unlock(&mountlist_mtx); return (mp); } } mtx_unlock(&mountlist_mtx); return ((struct mount *) 0); } /* * Get a new unique fsid. Try to make its val[0] unique, since this value * will be used to create fake device numbers for stat(). Also try (but * not so hard) make its val[0] unique mod 2^16, since some emulators only * support 16-bit device numbers. We end up with unique val[0]'s for the * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. * * Keep in mind that several mounts may be running in parallel. Starting * the search one past where the previous search terminated is both a * micro-optimization and a defense against returning the same fsid to * different mounts. */ void vfs_getnewfsid(mp) struct mount *mp; { static u_int16_t mntid_base; fsid_t tfsid; int mtype; mtx_lock(&mntid_mtx); mtype = mp->mnt_vfc->vfc_typenum; tfsid.val[1] = mtype; mtype = (mtype & 0xFF) << 24; for (;;) { tfsid.val[0] = makeudev(255, mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); mntid_base++; if (vfs_getvfs(&tfsid) == NULL) break; } mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; mtx_unlock(&mntid_mtx); } /* * Knob to control the precision of file timestamps: * * 0 = seconds only; nanoseconds zeroed. * 1 = seconds and nanoseconds, accurate within 1/HZ. * 2 = seconds and nanoseconds, truncated to microseconds. * >=3 = seconds and nanoseconds, maximum precision. */ enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; static int timestamp_precision = TSP_SEC; SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, ×tamp_precision, 0, ""); /* * Get a current timestamp. */ void vfs_timestamp(tsp) struct timespec *tsp; { struct timeval tv; switch (timestamp_precision) { case TSP_SEC: tsp->tv_sec = time_second; tsp->tv_nsec = 0; break; case TSP_HZ: getnanotime(tsp); break; case TSP_USEC: microtime(&tv); TIMEVAL_TO_TIMESPEC(&tv, tsp); break; case TSP_NSEC: default: nanotime(tsp); break; } } /* * Set vnode attributes to VNOVAL */ void vattr_null(vap) register struct vattr *vap; { vap->va_type = VNON; vap->va_size = VNOVAL; vap->va_bytes = VNOVAL; vap->va_mode = VNOVAL; vap->va_nlink = VNOVAL; vap->va_uid = VNOVAL; vap->va_gid = VNOVAL; vap->va_fsid = VNOVAL; vap->va_fileid = VNOVAL; vap->va_blocksize = VNOVAL; vap->va_rdev = VNOVAL; vap->va_atime.tv_sec = VNOVAL; vap->va_atime.tv_nsec = VNOVAL; vap->va_mtime.tv_sec = VNOVAL; vap->va_mtime.tv_nsec = VNOVAL; vap->va_ctime.tv_sec = VNOVAL; vap->va_ctime.tv_nsec = VNOVAL; vap->va_flags = VNOVAL; vap->va_gen = VNOVAL; vap->va_vaflags = 0; } /* * This routine is called when we have too many vnodes. It attempts * to free vnodes and will potentially free vnodes that still * have VM backing store (VM backing store is typically the cause * of a vnode blowout so we want to do this). Therefore, this operation * is not considered cheap. * * A number of conditions may prevent a vnode from being reclaimed. * the buffer cache may have references on the vnode, a directory * vnode may still have references due to the namei cache representing * underlying files, or the vnode may be in active use. It is not * desireable to reuse such vnodes. These conditions may cause the * number of vnodes to reach some minimum value regardless of what * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. */ static int vlrureclaim(struct mount *mp, int count) { struct vnode *vp; int done; int trigger; int usevnodes; /* * Calculate the trigger point, don't allow user * screwups to blow us up. This prevents us from * recycling vnodes with lots of resident pages. We * aren't trying to free memory, we are trying to * free vnodes. */ usevnodes = desiredvnodes; if (usevnodes <= 0) usevnodes = 1; trigger = cnt.v_page_count * 2 / usevnodes; done = 0; mtx_lock(&mntvnode_mtx); while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); if (vp->v_type != VNON && vp->v_type != VBAD && VMIGHTFREE(vp) && /* critical path opt */ (vp->v_object == NULL || vp->v_object->resident_page_count < trigger) && mtx_trylock(&vp->v_interlock) ) { mtx_unlock(&mntvnode_mtx); if (VMIGHTFREE(vp)) { vgonel(vp, curthread); done++; } else { mtx_unlock(&vp->v_interlock); } mtx_lock(&mntvnode_mtx); } --count; } mtx_unlock(&mntvnode_mtx); return done; } /* * Attempt to recycle vnodes in a context that is always safe to block. * Calling vlrurecycle() from the bowels of filesystem code has some * interesting deadlock problems. */ static struct proc *vnlruproc; static int vnlruproc_sig; static void vnlru_proc(void) { struct mount *mp, *nmp; int s; int done; struct proc *p = vnlruproc; struct thread *td = FIRST_THREAD_IN_PROC(p); /* XXXKSE */ mtx_lock(&Giant); EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, SHUTDOWN_PRI_FIRST); s = splbio(); for (;;) { kthread_suspend_check(p); if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { vnlruproc_sig = 0; tsleep(vnlruproc, PVFS, "vlruwt", 0); continue; } done = 0; mtx_lock(&mountlist_mtx); for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { nmp = TAILQ_NEXT(mp, mnt_list); continue; } done += vlrureclaim(mp, 10); mtx_lock(&mountlist_mtx); nmp = TAILQ_NEXT(mp, mnt_list); vfs_unbusy(mp, td); } mtx_unlock(&mountlist_mtx); if (done == 0) { #if 0 /* These messages are temporary debugging aids */ if (vnlru_nowhere < 5) printf("vnlru process getting nowhere..\n"); else if (vnlru_nowhere == 5) printf("vnlru process messages stopped.\n"); #endif vnlru_nowhere++; tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); } } splx(s); } static struct kproc_desc vnlru_kp = { "vnlru", vnlru_proc, &vnlruproc }; SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) /* * Routines having to do with the management of the vnode table. */ /* * Return the next vnode from the free list. */ int getnewvnode(tag, mp, vops, vpp) enum vtagtype tag; struct mount *mp; vop_t **vops; struct vnode **vpp; { int s; struct thread *td = curthread; /* XXX */ struct vnode *vp = NULL; struct mount *vnmp; vm_object_t object; s = splbio(); /* * Try to reuse vnodes if we hit the max. This situation only * occurs in certain large-memory (2G+) situations. We cannot * attempt to directly reclaim vnodes due to nasty recursion * problems. */ if (vnlruproc_sig == 0 && numvnodes - freevnodes > desiredvnodes) { vnlruproc_sig = 1; /* avoid unnecessary wakeups */ wakeup(vnlruproc); } /* * Attempt to reuse a vnode already on the free list, allocating * a new vnode if we can't find one or if we have not reached a * good minimum for good LRU performance. */ mtx_lock(&vnode_free_list_mtx); if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) { int count; for (count = 0; count < freevnodes; count++) { vp = TAILQ_FIRST(&vnode_free_list); if (vp == NULL || vp->v_usecount) panic("getnewvnode: free vnode isn't"); TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); /* Don't recycle if we can't get the interlock */ if (!mtx_trylock(&vp->v_interlock)) { vp = NULL; continue; } /* We should be able to immediately acquire this */ if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td) != 0) continue; /* * Don't recycle if we still have cached pages. */ if (VOP_GETVOBJECT(vp, &object) == 0 && (object->resident_page_count || object->ref_count)) { TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); VOP_UNLOCK(vp, 0, td); vp = NULL; continue; } if (LIST_FIRST(&vp->v_cache_src)) { /* * note: nameileafonly sysctl is temporary, * for debugging only, and will eventually be * removed. */ if (nameileafonly > 0) { /* * Do not reuse namei-cached directory * vnodes that have cached * subdirectories. */ if (cache_leaf_test(vp) < 0) { VOP_UNLOCK(vp, 0, td); TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); vp = NULL; continue; } } else if (nameileafonly < 0 || vmiodirenable == 0) { /* * Do not reuse namei-cached directory * vnodes if nameileafonly is -1 or * if VMIO backing for directories is * turned off (otherwise we reuse them * too quickly). */ VOP_UNLOCK(vp, 0, td); TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); vp = NULL; continue; } } /* * Skip over it if its filesystem is being suspended. */ if (vn_start_write(vp, &vnmp, V_NOWAIT) == 0) break; VOP_UNLOCK(vp, 0, td); TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); vp = NULL; } } if (vp) { vp->v_flag |= VDOOMED; vp->v_flag &= ~VFREE; freevnodes--; mtx_unlock(&vnode_free_list_mtx); cache_purge(vp); if (vp->v_type != VBAD) { VOP_UNLOCK(vp, 0, td); vgone(vp); } else { VOP_UNLOCK(vp, 0, td); } vn_finished_write(vnmp); #ifdef INVARIANTS { int s; if (vp->v_data) panic("cleaned vnode isn't"); s = splbio(); if (vp->v_numoutput) panic("Clean vnode has pending I/O's"); splx(s); if (vp->v_writecount != 0) panic("Non-zero write count"); } #endif if (vp->v_pollinfo) { mtx_destroy(&vp->v_pollinfo->vpi_lock); uma_zfree(vnodepoll_zone, vp->v_pollinfo); } vp->v_pollinfo = NULL; vp->v_flag = 0; vp->v_lastw = 0; vp->v_lasta = 0; vp->v_cstart = 0; vp->v_clen = 0; vp->v_socket = 0; } else { mtx_unlock(&vnode_free_list_mtx); vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK); bzero((char *) vp, sizeof *vp); mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); vp->v_dd = vp; cache_purge(vp); LIST_INIT(&vp->v_cache_src); TAILQ_INIT(&vp->v_cache_dst); numvnodes++; } TAILQ_INIT(&vp->v_cleanblkhd); TAILQ_INIT(&vp->v_dirtyblkhd); vp->v_type = VNON; vp->v_tag = tag; vp->v_op = vops; lockinit(&vp->v_lock, PVFS, "vnlock", VLKTIMEOUT, LK_NOPAUSE); insmntque(vp, mp); *vpp = vp; vp->v_usecount = 1; vp->v_data = 0; splx(s); #if 0 vnodeallocs++; if (vnodeallocs % vnoderecycleperiod == 0 && freevnodes < vnoderecycleminfreevn && vnoderecyclemintotalvn < numvnodes) { /* Recycle vnodes. */ cache_purgeleafdirs(vnoderecyclenumber); } #endif return (0); } /* * Move a vnode from one mount queue to another. */ static void insmntque(vp, mp) register struct vnode *vp; register struct mount *mp; { mtx_lock(&mntvnode_mtx); /* * Delete from old mount point vnode list, if on one. */ if (vp->v_mount != NULL) TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); /* * Insert into list of vnodes for the new mount point, if available. */ if ((vp->v_mount = mp) == NULL) { mtx_unlock(&mntvnode_mtx); return; } TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); mtx_unlock(&mntvnode_mtx); } /* * Update outstanding I/O count and do wakeup if requested. */ void vwakeup(bp) register struct buf *bp; { register struct vnode *vp; bp->b_flags &= ~B_WRITEINPROG; if ((vp = bp->b_vp)) { vp->v_numoutput--; if (vp->v_numoutput < 0) panic("vwakeup: neg numoutput"); if ((vp->v_numoutput == 0) && (vp->v_flag & VBWAIT)) { vp->v_flag &= ~VBWAIT; wakeup(&vp->v_numoutput); } } } /* * Flush out and invalidate all buffers associated with a vnode. * Called with the underlying object locked. */ int vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) register struct vnode *vp; int flags; struct ucred *cred; struct thread *td; int slpflag, slptimeo; { register struct buf *bp; struct buf *nbp, *blist; int s, error; vm_object_t object; GIANT_REQUIRED; if (flags & V_SAVE) { s = splbio(); while (vp->v_numoutput) { vp->v_flag |= VBWAIT; error = tsleep(&vp->v_numoutput, slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo); if (error) { splx(s); return (error); } } if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) { splx(s); if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) return (error); s = splbio(); if (vp->v_numoutput > 0 || !TAILQ_EMPTY(&vp->v_dirtyblkhd)) panic("vinvalbuf: dirty bufs"); } splx(s); } s = splbio(); for (;;) { blist = TAILQ_FIRST(&vp->v_cleanblkhd); if (!blist) blist = TAILQ_FIRST(&vp->v_dirtyblkhd); if (!blist) break; for (bp = blist; bp; bp = nbp) { nbp = TAILQ_NEXT(bp, b_vnbufs); if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, "vinvalbuf", slpflag, slptimeo); if (error == ENOLCK) break; splx(s); return (error); } /* * XXX Since there are no node locks for NFS, I * believe there is a slight chance that a delayed * write will occur while sleeping just above, so * check for it. Note that vfs_bio_awrite expects * buffers to reside on a queue, while BUF_WRITE and * brelse do not. */ if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && (flags & V_SAVE)) { if (bp->b_vp == vp) { if (bp->b_flags & B_CLUSTEROK) { BUF_UNLOCK(bp); vfs_bio_awrite(bp); } else { bremfree(bp); bp->b_flags |= B_ASYNC; BUF_WRITE(bp); } } else { bremfree(bp); (void) BUF_WRITE(bp); } break; } bremfree(bp); bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); bp->b_flags &= ~B_ASYNC; brelse(bp); } } /* * Wait for I/O to complete. XXX needs cleaning up. The vnode can * have write I/O in-progress but if there is a VM object then the * VM object can also have read-I/O in-progress. */ do { while (vp->v_numoutput > 0) { vp->v_flag |= VBWAIT; tsleep(&vp->v_numoutput, PVM, "vnvlbv", 0); } if (VOP_GETVOBJECT(vp, &object) == 0) { while (object->paging_in_progress) vm_object_pip_sleep(object, "vnvlbx"); } } while (vp->v_numoutput > 0); splx(s); /* * Destroy the copy in the VM cache, too. */ mtx_lock(&vp->v_interlock); if (VOP_GETVOBJECT(vp, &object) == 0) { vm_object_page_remove(object, 0, 0, (flags & V_SAVE) ? TRUE : FALSE); } mtx_unlock(&vp->v_interlock); if (!TAILQ_EMPTY(&vp->v_dirtyblkhd) || !TAILQ_EMPTY(&vp->v_cleanblkhd)) panic("vinvalbuf: flush failed"); return (0); } /* * Truncate a file's buffer and pages to a specified length. This * is in lieu of the old vinvalbuf mechanism, which performed unneeded * sync activity. */ int vtruncbuf(vp, cred, td, length, blksize) register struct vnode *vp; struct ucred *cred; struct thread *td; off_t length; int blksize; { register struct buf *bp; struct buf *nbp; int s, anyfreed; int trunclbn; /* * Round up to the *next* lbn. */ trunclbn = (length + blksize - 1) / blksize; s = splbio(); restart: anyfreed = 1; for (;anyfreed;) { anyfreed = 0; for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) { nbp = TAILQ_NEXT(bp, b_vnbufs); if (bp->b_lblkno >= trunclbn) { if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL); goto restart; } else { bremfree(bp); bp->b_flags |= (B_INVAL | B_RELBUF); bp->b_flags &= ~B_ASYNC; brelse(bp); anyfreed = 1; } if (nbp && (((nbp->b_xflags & BX_VNCLEAN) == 0) || (nbp->b_vp != vp) || (nbp->b_flags & B_DELWRI))) { goto restart; } } } for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { nbp = TAILQ_NEXT(bp, b_vnbufs); if (bp->b_lblkno >= trunclbn) { if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL); goto restart; } else { bremfree(bp); bp->b_flags |= (B_INVAL | B_RELBUF); bp->b_flags &= ~B_ASYNC; brelse(bp); anyfreed = 1; } if (nbp && (((nbp->b_xflags & BX_VNDIRTY) == 0) || (nbp->b_vp != vp) || (nbp->b_flags & B_DELWRI) == 0)) { goto restart; } } } } if (length > 0) { restartsync: for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { nbp = TAILQ_NEXT(bp, b_vnbufs); if ((bp->b_flags & B_DELWRI) && (bp->b_lblkno < 0)) { if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL); goto restart; } else { bremfree(bp); if (bp->b_vp == vp) { bp->b_flags |= B_ASYNC; } else { bp->b_flags &= ~B_ASYNC; } BUF_WRITE(bp); } goto restartsync; } } } while (vp->v_numoutput > 0) { vp->v_flag |= VBWAIT; tsleep(&vp->v_numoutput, PVM, "vbtrunc", 0); } splx(s); vnode_pager_setsize(vp, length); return (0); } /* * Associate a buffer with a vnode. */ void bgetvp(vp, bp) register struct vnode *vp; register struct buf *bp; { int s; KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); vhold(vp); bp->b_vp = vp; bp->b_dev = vn_todev(vp); /* * Insert onto list for new vnode. */ s = splbio(); bp->b_xflags |= BX_VNCLEAN; bp->b_xflags &= ~BX_VNDIRTY; TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs); splx(s); } /* * Disassociate a buffer from a vnode. */ void brelvp(bp) register struct buf *bp; { struct vnode *vp; struct buflists *listheadp; int s; KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); /* * Delete from old vnode list, if on one. */ vp = bp->b_vp; s = splbio(); if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) { if (bp->b_xflags & BX_VNDIRTY) listheadp = &vp->v_dirtyblkhd; else listheadp = &vp->v_cleanblkhd; TAILQ_REMOVE(listheadp, bp, b_vnbufs); bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); } if ((vp->v_flag & VONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) { vp->v_flag &= ~VONWORKLST; LIST_REMOVE(vp, v_synclist); } splx(s); bp->b_vp = (struct vnode *) 0; vdrop(vp); - if (bp->b_object) { - vm_object_deallocate(bp->b_object); + if (bp->b_object) bp->b_object = NULL; - } } /* * Add an item to the syncer work queue. */ static void vn_syncer_add_to_worklist(struct vnode *vp, int delay) { int s, slot; s = splbio(); if (vp->v_flag & VONWORKLST) { LIST_REMOVE(vp, v_synclist); } if (delay > syncer_maxdelay - 2) delay = syncer_maxdelay - 2; slot = (syncer_delayno + delay) & syncer_mask; LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); vp->v_flag |= VONWORKLST; splx(s); } struct proc *updateproc; static void sched_sync(void); static struct kproc_desc up_kp = { "syncer", sched_sync, &updateproc }; SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) /* * System filesystem synchronizer daemon. */ void sched_sync(void) { struct synclist *slp; struct vnode *vp; struct mount *mp; long starttime; int s; struct thread *td = FIRST_THREAD_IN_PROC(updateproc); /* XXXKSE */ mtx_lock(&Giant); EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, td->td_proc, SHUTDOWN_PRI_LAST); for (;;) { kthread_suspend_check(td->td_proc); starttime = time_second; /* * Push files whose dirty time has expired. Be careful * of interrupt race on slp queue. */ s = splbio(); slp = &syncer_workitem_pending[syncer_delayno]; syncer_delayno += 1; if (syncer_delayno == syncer_maxdelay) syncer_delayno = 0; splx(s); while ((vp = LIST_FIRST(slp)) != NULL) { if (VOP_ISLOCKED(vp, NULL) == 0 && vn_start_write(vp, &mp, V_NOWAIT) == 0) { vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); (void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td); VOP_UNLOCK(vp, 0, td); vn_finished_write(mp); } s = splbio(); if (LIST_FIRST(slp) == vp) { /* * Note: v_tag VT_VFS vps can remain on the * worklist too with no dirty blocks, but * since sync_fsync() moves it to a different * slot we are safe. */ if (TAILQ_EMPTY(&vp->v_dirtyblkhd) && !vn_isdisk(vp, NULL)) panic("sched_sync: fsync failed vp %p tag %d", vp, vp->v_tag); /* * Put us back on the worklist. The worklist * routine will remove us from our current * position and then add us back in at a later * position. */ vn_syncer_add_to_worklist(vp, syncdelay); } splx(s); } /* * Do soft update processing. */ if (softdep_process_worklist_hook != NULL) (*softdep_process_worklist_hook)(NULL); /* * The variable rushjob allows the kernel to speed up the * processing of the filesystem syncer process. A rushjob * value of N tells the filesystem syncer to process the next * N seconds worth of work on its queue ASAP. Currently rushjob * is used by the soft update code to speed up the filesystem * syncer process when the incore state is getting so far * ahead of the disk that the kernel memory pool is being * threatened with exhaustion. */ if (rushjob > 0) { rushjob -= 1; continue; } /* * If it has taken us less than a second to process the * current work, then wait. Otherwise start right over * again. We can still lose time if any single round * takes more than two seconds, but it does not really * matter as we are just trying to generally pace the * filesystem activity. */ if (time_second == starttime) tsleep(&lbolt, PPAUSE, "syncer", 0); } } /* * Request the syncer daemon to speed up its work. * We never push it to speed up more than half of its * normal turn time, otherwise it could take over the cpu. * XXXKSE only one update? */ int speedup_syncer() { mtx_lock_spin(&sched_lock); if (FIRST_THREAD_IN_PROC(updateproc)->td_wchan == &lbolt) /* XXXKSE */ setrunnable(FIRST_THREAD_IN_PROC(updateproc)); mtx_unlock_spin(&sched_lock); if (rushjob < syncdelay / 2) { rushjob += 1; stat_rush_requests += 1; return (1); } return(0); } /* * Associate a p-buffer with a vnode. * * Also sets B_PAGING flag to indicate that vnode is not fully associated * with the buffer. i.e. the bp has not been linked into the vnode or * ref-counted. */ void pbgetvp(vp, bp) register struct vnode *vp; register struct buf *bp; { KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); bp->b_vp = vp; bp->b_flags |= B_PAGING; bp->b_dev = vn_todev(vp); } /* * Disassociate a p-buffer from a vnode. */ void pbrelvp(bp) register struct buf *bp; { KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); /* XXX REMOVE ME */ if (TAILQ_NEXT(bp, b_vnbufs) != NULL) { panic( "relpbuf(): b_vp was probably reassignbuf()d %p %x", bp, (int)bp->b_flags ); } bp->b_vp = (struct vnode *) 0; bp->b_flags &= ~B_PAGING; } /* * Reassign a buffer from one vnode to another. * Used to assign file specific control information * (indirect blocks) to the vnode to which they belong. */ void reassignbuf(bp, newvp) register struct buf *bp; register struct vnode *newvp; { struct buflists *listheadp; int delay; int s; if (newvp == NULL) { printf("reassignbuf: NULL"); return; } ++reassignbufcalls; /* * B_PAGING flagged buffers cannot be reassigned because their vp * is not fully linked in. */ if (bp->b_flags & B_PAGING) panic("cannot reassign paging buffer"); s = splbio(); /* * Delete from old vnode list, if on one. */ if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) { if (bp->b_xflags & BX_VNDIRTY) listheadp = &bp->b_vp->v_dirtyblkhd; else listheadp = &bp->b_vp->v_cleanblkhd; TAILQ_REMOVE(listheadp, bp, b_vnbufs); bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); if (bp->b_vp != newvp) { vdrop(bp->b_vp); bp->b_vp = NULL; /* for clarification */ } } /* * If dirty, put on list of dirty buffers; otherwise insert onto list * of clean buffers. */ if (bp->b_flags & B_DELWRI) { struct buf *tbp; listheadp = &newvp->v_dirtyblkhd; if ((newvp->v_flag & VONWORKLST) == 0) { switch (newvp->v_type) { case VDIR: delay = dirdelay; break; case VCHR: if (newvp->v_rdev->si_mountpoint != NULL) { delay = metadelay; break; } /* fall through */ default: delay = filedelay; } vn_syncer_add_to_worklist(newvp, delay); } bp->b_xflags |= BX_VNDIRTY; tbp = TAILQ_FIRST(listheadp); if (tbp == NULL || bp->b_lblkno == 0 || (bp->b_lblkno > 0 && tbp->b_lblkno < 0) || (bp->b_lblkno > 0 && bp->b_lblkno < tbp->b_lblkno)) { TAILQ_INSERT_HEAD(listheadp, bp, b_vnbufs); ++reassignbufsortgood; } else if (bp->b_lblkno < 0) { TAILQ_INSERT_TAIL(listheadp, bp, b_vnbufs); ++reassignbufsortgood; } else if (reassignbufmethod == 1) { /* * New sorting algorithm, only handle sequential case, * otherwise append to end (but before metadata) */ if ((tbp = gbincore(newvp, bp->b_lblkno - 1)) != NULL && (tbp->b_xflags & BX_VNDIRTY)) { /* * Found the best place to insert the buffer */ TAILQ_INSERT_AFTER(listheadp, tbp, bp, b_vnbufs); ++reassignbufsortgood; } else { /* * Missed, append to end, but before meta-data. * We know that the head buffer in the list is * not meta-data due to prior conditionals. * * Indirect effects: NFS second stage write * tends to wind up here, giving maximum * distance between the unstable write and the * commit rpc. */ tbp = TAILQ_LAST(listheadp, buflists); while (tbp && tbp->b_lblkno < 0) tbp = TAILQ_PREV(tbp, buflists, b_vnbufs); TAILQ_INSERT_AFTER(listheadp, tbp, bp, b_vnbufs); ++reassignbufsortbad; } } else { /* * Old sorting algorithm, scan queue and insert */ struct buf *ttbp; while ((ttbp = TAILQ_NEXT(tbp, b_vnbufs)) && (ttbp->b_lblkno < bp->b_lblkno)) { ++reassignbufloops; tbp = ttbp; } TAILQ_INSERT_AFTER(listheadp, tbp, bp, b_vnbufs); } } else { bp->b_xflags |= BX_VNCLEAN; TAILQ_INSERT_TAIL(&newvp->v_cleanblkhd, bp, b_vnbufs); if ((newvp->v_flag & VONWORKLST) && TAILQ_EMPTY(&newvp->v_dirtyblkhd)) { newvp->v_flag &= ~VONWORKLST; LIST_REMOVE(newvp, v_synclist); } } if (bp->b_vp != newvp) { bp->b_vp = newvp; vhold(bp->b_vp); } splx(s); } /* * Create a vnode for a device. * Used for mounting the root filesystem. */ int bdevvp(dev, vpp) dev_t dev; struct vnode **vpp; { register struct vnode *vp; struct vnode *nvp; int error; if (dev == NODEV) { *vpp = NULLVP; return (ENXIO); } if (vfinddev(dev, VCHR, vpp)) return (0); error = getnewvnode(VT_NON, (struct mount *)0, spec_vnodeop_p, &nvp); if (error) { *vpp = NULLVP; return (error); } vp = nvp; vp->v_type = VCHR; addalias(vp, dev); *vpp = vp; return (0); } /* * Add vnode to the alias list hung off the dev_t. * * The reason for this gunk is that multiple vnodes can reference * the same physical device, so checking vp->v_usecount to see * how many users there are is inadequate; the v_usecount for * the vnodes need to be accumulated. vcount() does that. */ struct vnode * addaliasu(nvp, nvp_rdev) struct vnode *nvp; udev_t nvp_rdev; { struct vnode *ovp; vop_t **ops; dev_t dev; if (nvp->v_type == VBLK) return (nvp); if (nvp->v_type != VCHR) panic("addaliasu on non-special vnode"); dev = udev2dev(nvp_rdev, 0); /* * Check to see if we have a bdevvp vnode with no associated * filesystem. If so, we want to associate the filesystem of * the new newly instigated vnode with the bdevvp vnode and * discard the newly created vnode rather than leaving the * bdevvp vnode lying around with no associated filesystem. */ if (vfinddev(dev, nvp->v_type, &ovp) == 0 || ovp->v_data != NULL) { addalias(nvp, dev); return (nvp); } /* * Discard unneeded vnode, but save its node specific data. * Note that if there is a lock, it is carried over in the * node specific data to the replacement vnode. */ vref(ovp); ovp->v_data = nvp->v_data; ovp->v_tag = nvp->v_tag; nvp->v_data = NULL; lockinit(&ovp->v_lock, PVFS, nvp->v_lock.lk_wmesg, nvp->v_lock.lk_timo, nvp->v_lock.lk_flags & LK_EXTFLG_MASK); if (nvp->v_vnlock) ovp->v_vnlock = &ovp->v_lock; ops = ovp->v_op; ovp->v_op = nvp->v_op; if (VOP_ISLOCKED(nvp, curthread)) { VOP_UNLOCK(nvp, 0, curthread); vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); } nvp->v_op = ops; insmntque(ovp, nvp->v_mount); vrele(nvp); vgone(nvp); return (ovp); } /* This is a local helper function that do the same as addaliasu, but for a * dev_t instead of an udev_t. */ static void addalias(nvp, dev) struct vnode *nvp; dev_t dev; { KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); nvp->v_rdev = dev; mtx_lock(&spechash_mtx); SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); mtx_unlock(&spechash_mtx); } /* * Grab a particular vnode from the free list, increment its * reference count and lock it. The vnode lock bit is set if the * vnode is being eliminated in vgone. The process is awakened * when the transition is completed, and an error returned to * indicate that the vnode is no longer usable (possibly having * been changed to a new filesystem type). */ int vget(vp, flags, td) register struct vnode *vp; int flags; struct thread *td; { int error; /* * If the vnode is in the process of being cleaned out for * another use, we wait for the cleaning to finish and then * return failure. Cleaning is determined by checking that * the VXLOCK flag is set. */ if ((flags & LK_INTERLOCK) == 0) mtx_lock(&vp->v_interlock); if (vp->v_flag & VXLOCK) { if (vp->v_vxproc == curthread) { #if 0 /* this can now occur in normal operation */ log(LOG_INFO, "VXLOCK interlock avoided\n"); #endif } else { vp->v_flag |= VXWANT; msleep(vp, &vp->v_interlock, PINOD | PDROP, "vget", 0); return (ENOENT); } } vp->v_usecount++; if (VSHOULDBUSY(vp)) vbusy(vp); if (flags & LK_TYPE_MASK) { if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { /* * must expand vrele here because we do not want * to call VOP_INACTIVE if the reference count * drops back to zero since it was never really * active. We must remove it from the free list * before sleeping so that multiple processes do * not try to recycle it. */ mtx_lock(&vp->v_interlock); vp->v_usecount--; if (VSHOULDFREE(vp)) vfree(vp); else vlruvp(vp); mtx_unlock(&vp->v_interlock); } return (error); } mtx_unlock(&vp->v_interlock); return (0); } /* * Increase the reference count of a vnode. */ void vref(struct vnode *vp) { mtx_lock(&vp->v_interlock); vp->v_usecount++; mtx_unlock(&vp->v_interlock); } /* * Vnode put/release. * If count drops to zero, call inactive routine and return to freelist. */ void vrele(vp) struct vnode *vp; { struct thread *td = curthread; /* XXX */ KASSERT(vp != NULL, ("vrele: null vp")); mtx_lock(&vp->v_interlock); /* Skip this v_writecount check if we're going to panic below. */ KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, ("vrele: missed vn_close")); if (vp->v_usecount > 1) { vp->v_usecount--; mtx_unlock(&vp->v_interlock); return; } if (vp->v_usecount == 1) { vp->v_usecount--; /* * We must call VOP_INACTIVE with the node locked. * If we are doing a vput, the node is already locked, * but, in the case of vrele, we must explicitly lock * the vnode before calling VOP_INACTIVE. */ if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) VOP_INACTIVE(vp, td); if (VSHOULDFREE(vp)) vfree(vp); else vlruvp(vp); } else { #ifdef DIAGNOSTIC vprint("vrele: negative ref count", vp); mtx_unlock(&vp->v_interlock); #endif panic("vrele: negative ref cnt"); } } /* * Release an already locked vnode. This give the same effects as * unlock+vrele(), but takes less time and avoids releasing and * re-aquiring the lock (as vrele() aquires the lock internally.) */ void vput(vp) struct vnode *vp; { struct thread *td = curthread; /* XXX */ GIANT_REQUIRED; KASSERT(vp != NULL, ("vput: null vp")); mtx_lock(&vp->v_interlock); /* Skip this v_writecount check if we're going to panic below. */ KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, ("vput: missed vn_close")); if (vp->v_usecount > 1) { vp->v_usecount--; VOP_UNLOCK(vp, LK_INTERLOCK, td); return; } if (vp->v_usecount == 1) { vp->v_usecount--; /* * We must call VOP_INACTIVE with the node locked. * If we are doing a vput, the node is already locked, * so we just need to release the vnode mutex. */ mtx_unlock(&vp->v_interlock); VOP_INACTIVE(vp, td); if (VSHOULDFREE(vp)) vfree(vp); else vlruvp(vp); } else { #ifdef DIAGNOSTIC vprint("vput: negative ref count", vp); #endif panic("vput: negative ref cnt"); } } /* * Somebody doesn't want the vnode recycled. */ void vhold(vp) register struct vnode *vp; { int s; s = splbio(); vp->v_holdcnt++; if (VSHOULDBUSY(vp)) vbusy(vp); splx(s); } /* * Note that there is one less who cares about this vnode. vdrop() is the * opposite of vhold(). */ void vdrop(vp) register struct vnode *vp; { int s; s = splbio(); if (vp->v_holdcnt <= 0) panic("vdrop: holdcnt"); vp->v_holdcnt--; if (VSHOULDFREE(vp)) vfree(vp); else vlruvp(vp); splx(s); } /* * Remove any vnodes in the vnode table belonging to mount point mp. * * If FORCECLOSE is not specified, there should not be any active ones, * return error if any are found (nb: this is a user error, not a * system error). If FORCECLOSE is specified, detach any active vnodes * that are found. * * If WRITECLOSE is set, only flush out regular file vnodes open for * writing. * * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped. * * `rootrefs' specifies the base reference count for the root vnode * of this filesystem. The root vnode is considered busy if its * v_usecount exceeds this value. On a successful return, vflush() * will call vrele() on the root vnode exactly rootrefs times. * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must * be zero. */ #ifdef DIAGNOSTIC static int busyprt = 0; /* print out busy vnodes */ SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); #endif int vflush(mp, rootrefs, flags) struct mount *mp; int rootrefs; int flags; { struct thread *td = curthread; /* XXX */ struct vnode *vp, *nvp, *rootvp = NULL; struct vattr vattr; int busy = 0, error; if (rootrefs > 0) { KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, ("vflush: bad args")); /* * Get the filesystem root vnode. We can vput() it * immediately, since with rootrefs > 0, it won't go away. */ if ((error = VFS_ROOT(mp, &rootvp)) != 0) return (error); vput(rootvp); } mtx_lock(&mntvnode_mtx); loop: for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp; vp = nvp) { /* * Make sure this vnode wasn't reclaimed in getnewvnode(). * Start over if it has (it won't be on the list anymore). */ if (vp->v_mount != mp) goto loop; nvp = TAILQ_NEXT(vp, v_nmntvnodes); mtx_unlock(&mntvnode_mtx); mtx_lock(&vp->v_interlock); /* * Skip over a vnodes marked VSYSTEM. */ if ((flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) { mtx_unlock(&vp->v_interlock); mtx_lock(&mntvnode_mtx); continue; } /* * If WRITECLOSE is set, flush out unlinked but still open * files (even if open only for reading) and regular file * vnodes open for writing. */ if ((flags & WRITECLOSE) && (vp->v_type == VNON || (VOP_GETATTR(vp, &vattr, td->td_ucred, td) == 0 && vattr.va_nlink > 0)) && (vp->v_writecount == 0 || vp->v_type != VREG)) { mtx_unlock(&vp->v_interlock); mtx_lock(&mntvnode_mtx); continue; } /* * With v_usecount == 0, all we need to do is clear out the * vnode data structures and we are done. */ if (vp->v_usecount == 0) { vgonel(vp, td); mtx_lock(&mntvnode_mtx); continue; } /* * If FORCECLOSE is set, forcibly close the vnode. For block * or character devices, revert to an anonymous device. For * all other files, just kill them. */ if (flags & FORCECLOSE) { if (vp->v_type != VCHR) { vgonel(vp, td); } else { vclean(vp, 0, td); vp->v_op = spec_vnodeop_p; insmntque(vp, (struct mount *) 0); } mtx_lock(&mntvnode_mtx); continue; } #ifdef DIAGNOSTIC if (busyprt) vprint("vflush: busy vnode", vp); #endif mtx_unlock(&vp->v_interlock); mtx_lock(&mntvnode_mtx); busy++; } mtx_unlock(&mntvnode_mtx); if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { /* * If just the root vnode is busy, and if its refcount * is equal to `rootrefs', then go ahead and kill it. */ mtx_lock(&rootvp->v_interlock); KASSERT(busy > 0, ("vflush: not busy")); KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs")); if (busy == 1 && rootvp->v_usecount == rootrefs) { vgonel(rootvp, td); busy = 0; } else mtx_unlock(&rootvp->v_interlock); } if (busy) return (EBUSY); for (; rootrefs > 0; rootrefs--) vrele(rootvp); return (0); } /* * This moves a now (likely recyclable) vnode to the end of the * mountlist. XXX However, it is temporarily disabled until we * can clean up ffs_sync() and friends, which have loop restart * conditions which this code causes to operate O(N^2). */ static void vlruvp(struct vnode *vp) { #if 0 struct mount *mp; if ((mp = vp->v_mount) != NULL) { mtx_lock(&mntvnode_mtx); TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); mtx_unlock(&mntvnode_mtx); } #endif } /* * Disassociate the underlying filesystem from a vnode. */ static void vclean(vp, flags, td) struct vnode *vp; int flags; struct thread *td; { int active; /* * Check to see if the vnode is in use. If so we have to reference it * before we clean it out so that its count cannot fall to zero and * generate a race against ourselves to recycle it. */ if ((active = vp->v_usecount)) vp->v_usecount++; /* * Prevent the vnode from being recycled or brought into use while we * clean it out. */ if (vp->v_flag & VXLOCK) panic("vclean: deadlock"); vp->v_flag |= VXLOCK; vp->v_vxproc = curthread; /* * Even if the count is zero, the VOP_INACTIVE routine may still * have the object locked while it cleans it out. The VOP_LOCK * ensures that the VOP_INACTIVE routine is done with its work. * For active vnodes, it ensures that no other activity can * occur while the underlying object is being cleaned out. */ VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); /* * Clean out any buffers associated with the vnode. * If the flush fails, just toss the buffers. */ if (flags & DOCLOSE) { if (TAILQ_FIRST(&vp->v_dirtyblkhd) != NULL) (void) vn_write_suspend_wait(vp, NULL, V_WAIT); if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) vinvalbuf(vp, 0, NOCRED, td, 0, 0); } VOP_DESTROYVOBJECT(vp); /* * If purging an active vnode, it must be closed and * deactivated before being reclaimed. Note that the * VOP_INACTIVE will unlock the vnode. */ if (active) { if (flags & DOCLOSE) VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); VOP_INACTIVE(vp, td); } else { /* * Any other processes trying to obtain this lock must first * wait for VXLOCK to clear, then call the new lock operation. */ VOP_UNLOCK(vp, 0, td); } /* * Reclaim the vnode. */ if (VOP_RECLAIM(vp, td)) panic("vclean: cannot reclaim"); if (active) { /* * Inline copy of vrele() since VOP_INACTIVE * has already been called. */ mtx_lock(&vp->v_interlock); if (--vp->v_usecount <= 0) { #ifdef DIAGNOSTIC if (vp->v_usecount < 0 || vp->v_writecount != 0) { vprint("vclean: bad ref count", vp); panic("vclean: ref cnt"); } #endif vfree(vp); } mtx_unlock(&vp->v_interlock); } cache_purge(vp); vp->v_vnlock = NULL; lockdestroy(&vp->v_lock); if (VSHOULDFREE(vp)) vfree(vp); /* * Done with purge, notify sleepers of the grim news. */ vp->v_op = dead_vnodeop_p; if (vp->v_pollinfo != NULL) vn_pollgone(vp); vp->v_tag = VT_NON; vp->v_flag &= ~VXLOCK; vp->v_vxproc = NULL; if (vp->v_flag & VXWANT) { vp->v_flag &= ~VXWANT; wakeup(vp); } } /* * Eliminate all activity associated with the requested vnode * and with all vnodes aliased to the requested vnode. */ int vop_revoke(ap) struct vop_revoke_args /* { struct vnode *a_vp; int a_flags; } */ *ap; { struct vnode *vp, *vq; dev_t dev; KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); vp = ap->a_vp; /* * If a vgone (or vclean) is already in progress, * wait until it is done and return. */ if (vp->v_flag & VXLOCK) { vp->v_flag |= VXWANT; msleep(vp, &vp->v_interlock, PINOD | PDROP, "vop_revokeall", 0); return (0); } dev = vp->v_rdev; for (;;) { mtx_lock(&spechash_mtx); vq = SLIST_FIRST(&dev->si_hlist); mtx_unlock(&spechash_mtx); if (!vq) break; vgone(vq); } return (0); } /* * Recycle an unused vnode to the front of the free list. * Release the passed interlock if the vnode will be recycled. */ int vrecycle(vp, inter_lkp, td) struct vnode *vp; struct mtx *inter_lkp; struct thread *td; { mtx_lock(&vp->v_interlock); if (vp->v_usecount == 0) { if (inter_lkp) { mtx_unlock(inter_lkp); } vgonel(vp, td); return (1); } mtx_unlock(&vp->v_interlock); return (0); } /* * Eliminate all activity associated with a vnode * in preparation for reuse. */ void vgone(vp) register struct vnode *vp; { struct thread *td = curthread; /* XXX */ mtx_lock(&vp->v_interlock); vgonel(vp, td); } /* * vgone, with the vp interlock held. */ void vgonel(vp, td) struct vnode *vp; struct thread *td; { int s; /* * If a vgone (or vclean) is already in progress, * wait until it is done and return. */ if (vp->v_flag & VXLOCK) { vp->v_flag |= VXWANT; msleep(vp, &vp->v_interlock, PINOD | PDROP, "vgone", 0); return; } /* * Clean out the filesystem specific data. */ vclean(vp, DOCLOSE, td); mtx_lock(&vp->v_interlock); /* * Delete from old mount point vnode list, if on one. */ if (vp->v_mount != NULL) insmntque(vp, (struct mount *)0); /* * If special device, remove it from special device alias list * if it is on one. */ if (vp->v_type == VCHR && vp->v_rdev != NULL && vp->v_rdev != NODEV) { mtx_lock(&spechash_mtx); SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext); freedev(vp->v_rdev); mtx_unlock(&spechash_mtx); vp->v_rdev = NULL; } /* * If it is on the freelist and not already at the head, * move it to the head of the list. The test of the * VDOOMED flag and the reference count of zero is because * it will be removed from the free list by getnewvnode, * but will not have its reference count incremented until * after calling vgone. If the reference count were * incremented first, vgone would (incorrectly) try to * close the previous instance of the underlying object. */ if (vp->v_usecount == 0 && !(vp->v_flag & VDOOMED)) { s = splbio(); mtx_lock(&vnode_free_list_mtx); if (vp->v_flag & VFREE) TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); else freevnodes++; vp->v_flag |= VFREE; TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); mtx_unlock(&vnode_free_list_mtx); splx(s); } vp->v_type = VBAD; mtx_unlock(&vp->v_interlock); } /* * Lookup a vnode by device number. */ int vfinddev(dev, type, vpp) dev_t dev; enum vtype type; struct vnode **vpp; { struct vnode *vp; mtx_lock(&spechash_mtx); SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) { if (type == vp->v_type) { *vpp = vp; mtx_unlock(&spechash_mtx); return (1); } } mtx_unlock(&spechash_mtx); return (0); } /* * Calculate the total number of references to a special device. */ int vcount(vp) struct vnode *vp; { struct vnode *vq; int count; count = 0; mtx_lock(&spechash_mtx); SLIST_FOREACH(vq, &vp->v_rdev->si_hlist, v_specnext) count += vq->v_usecount; mtx_unlock(&spechash_mtx); return (count); } /* * Same as above, but using the dev_t as argument */ int count_dev(dev) dev_t dev; { struct vnode *vp; vp = SLIST_FIRST(&dev->si_hlist); if (vp == NULL) return (0); return(vcount(vp)); } /* * Print out a description of a vnode. */ static char *typename[] = {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; void vprint(label, vp) char *label; struct vnode *vp; { char buf[96]; if (label != NULL) printf("%s: %p: ", label, (void *)vp); else printf("%p: ", (void *)vp); printf("type %s, usecount %d, writecount %d, refcount %d,", typename[vp->v_type], vp->v_usecount, vp->v_writecount, vp->v_holdcnt); buf[0] = '\0'; if (vp->v_flag & VROOT) strcat(buf, "|VROOT"); if (vp->v_flag & VTEXT) strcat(buf, "|VTEXT"); if (vp->v_flag & VSYSTEM) strcat(buf, "|VSYSTEM"); if (vp->v_flag & VXLOCK) strcat(buf, "|VXLOCK"); if (vp->v_flag & VXWANT) strcat(buf, "|VXWANT"); if (vp->v_flag & VBWAIT) strcat(buf, "|VBWAIT"); if (vp->v_flag & VDOOMED) strcat(buf, "|VDOOMED"); if (vp->v_flag & VFREE) strcat(buf, "|VFREE"); if (vp->v_flag & VOBJBUF) strcat(buf, "|VOBJBUF"); if (buf[0] != '\0') printf(" flags (%s)", &buf[1]); if (vp->v_data == NULL) { printf("\n"); } else { printf("\n\t"); VOP_PRINT(vp); } } #ifdef DDB #include /* * List all of the locked vnodes in the system. * Called when debugging the kernel. */ DB_SHOW_COMMAND(lockedvnods, lockedvnodes) { struct thread *td = curthread; /* XXX */ struct mount *mp, *nmp; struct vnode *vp; printf("Locked vnodes\n"); mtx_lock(&mountlist_mtx); for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { nmp = TAILQ_NEXT(mp, mnt_list); continue; } mtx_lock(&mntvnode_mtx); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (VOP_ISLOCKED(vp, NULL)) vprint((char *)0, vp); } mtx_unlock(&mntvnode_mtx); mtx_lock(&mountlist_mtx); nmp = TAILQ_NEXT(mp, mnt_list); vfs_unbusy(mp, td); } mtx_unlock(&mountlist_mtx); } #endif /* * Top level filesystem related information gathering. */ static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); static int vfs_sysctl(SYSCTL_HANDLER_ARGS) { int *name = (int *)arg1 - 1; /* XXX */ u_int namelen = arg2 + 1; /* XXX */ struct vfsconf *vfsp; #if 1 || defined(COMPAT_PRELITE2) /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ if (namelen == 1) return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); #endif /* XXX the below code does not compile; vfs_sysctl does not exist. */ #ifdef notyet /* all sysctl names at this level are at least name and field */ if (namelen < 2) return (ENOTDIR); /* overloaded */ if (name[0] != VFS_GENERIC) { for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) if (vfsp->vfc_typenum == name[0]) break; if (vfsp == NULL) return (EOPNOTSUPP); return ((*vfsp->vfc_vfsops->vfs_sysctl)(&name[1], namelen - 1, oldp, oldlenp, newp, newlen, td)); } #endif switch (name[1]) { case VFS_MAXTYPENUM: if (namelen != 2) return (ENOTDIR); return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); case VFS_CONF: if (namelen != 3) return (ENOTDIR); /* overloaded */ for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) if (vfsp->vfc_typenum == name[2]) break; if (vfsp == NULL) return (EOPNOTSUPP); return (SYSCTL_OUT(req, vfsp, sizeof *vfsp)); } return (EOPNOTSUPP); } SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD, vfs_sysctl, "Generic filesystem"); #if 1 || defined(COMPAT_PRELITE2) static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) { int error; struct vfsconf *vfsp; struct ovfsconf ovfs; for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ strcpy(ovfs.vfc_name, vfsp->vfc_name); ovfs.vfc_index = vfsp->vfc_typenum; ovfs.vfc_refcount = vfsp->vfc_refcount; ovfs.vfc_flags = vfsp->vfc_flags; error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); if (error) return error; } return 0; } #endif /* 1 || COMPAT_PRELITE2 */ #if COMPILING_LINT #define KINFO_VNODESLOP 10 /* * Dump vnode list (via sysctl). * Copyout address of vnode followed by vnode. */ /* ARGSUSED */ static int sysctl_vnode(SYSCTL_HANDLER_ARGS) { struct thread *td = curthread; /* XXX */ struct mount *mp, *nmp; struct vnode *nvp, *vp; int error; #define VPTRSZ sizeof (struct vnode *) #define VNODESZ sizeof (struct vnode) req->lock = 0; if (!req->oldptr) /* Make an estimate */ return (SYSCTL_OUT(req, 0, (numvnodes + KINFO_VNODESLOP) * (VPTRSZ + VNODESZ))); mtx_lock(&mountlist_mtx); for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { nmp = TAILQ_NEXT(mp, mnt_list); continue; } mtx_lock(&mntvnode_mtx); again: for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) { /* * Check that the vp is still associated with * this filesystem. RACE: could have been * recycled onto the same filesystem. */ if (vp->v_mount != mp) goto again; nvp = TAILQ_NEXT(vp, v_nmntvnodes); mtx_unlock(&mntvnode_mtx); if ((error = SYSCTL_OUT(req, &vp, VPTRSZ)) || (error = SYSCTL_OUT(req, vp, VNODESZ))) return (error); mtx_lock(&mntvnode_mtx); } mtx_unlock(&mntvnode_mtx); mtx_lock(&mountlist_mtx); nmp = TAILQ_NEXT(mp, mnt_list); vfs_unbusy(mp, td); } mtx_unlock(&mountlist_mtx); return (0); } /* * XXX * Exporting the vnode list on large systems causes them to crash. * Exporting the vnode list on medium systems causes sysctl to coredump. */ SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 0, 0, sysctl_vnode, "S,vnode", ""); #endif /* * Check to see if a filesystem is mounted on a block device. */ int vfs_mountedon(vp) struct vnode *vp; { if (vp->v_rdev->si_mountpoint != NULL) return (EBUSY); return (0); } /* * Unmount all filesystems. The list is traversed in reverse order * of mounting to avoid dependencies. */ void vfs_unmountall() { struct mount *mp; struct thread *td; int error; if (curthread != NULL) td = curthread; else td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ /* * Since this only runs when rebooting, it is not interlocked. */ while(!TAILQ_EMPTY(&mountlist)) { mp = TAILQ_LAST(&mountlist, mntlist); error = dounmount(mp, MNT_FORCE, td); if (error) { TAILQ_REMOVE(&mountlist, mp, mnt_list); printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); if (error == EBUSY) printf("BUSY)\n"); else printf("%d)\n", error); } else { /* The unmount has removed mp from the mountlist */ } } } /* * perform msync on all vnodes under a mount point * the mount point must be locked. */ void vfs_msync(struct mount *mp, int flags) { struct vnode *vp, *nvp; struct vm_object *obj; int tries; GIANT_REQUIRED; tries = 5; mtx_lock(&mntvnode_mtx); loop: for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) { if (vp->v_mount != mp) { if (--tries > 0) goto loop; break; } nvp = TAILQ_NEXT(vp, v_nmntvnodes); if (vp->v_flag & VXLOCK) /* XXX: what if MNT_WAIT? */ continue; if (vp->v_flag & VNOSYNC) /* unlinked, skip it */ continue; if ((vp->v_flag & VOBJDIRTY) && (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { mtx_unlock(&mntvnode_mtx); if (!vget(vp, LK_EXCLUSIVE | LK_RETRY | LK_NOOBJ, curthread)) { if (VOP_GETVOBJECT(vp, &obj) == 0) { vm_object_page_clean(obj, 0, 0, flags == MNT_WAIT ? OBJPC_SYNC : OBJPC_NOSYNC); } vput(vp); } mtx_lock(&mntvnode_mtx); if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { if (--tries > 0) goto loop; break; } } } mtx_unlock(&mntvnode_mtx); } /* * Create the VM object needed for VMIO and mmap support. This * is done for all VREG files in the system. Some filesystems might * afford the additional metadata buffering capability of the * VMIO code by making the device node be VMIO mode also. * * vp must be locked when vfs_object_create is called. */ int vfs_object_create(vp, td, cred) struct vnode *vp; struct thread *td; struct ucred *cred; { GIANT_REQUIRED; return (VOP_CREATEVOBJECT(vp, cred, td)); } /* * Mark a vnode as free, putting it up for recycling. */ void vfree(vp) struct vnode *vp; { int s; s = splbio(); mtx_lock(&vnode_free_list_mtx); KASSERT((vp->v_flag & VFREE) == 0, ("vnode already free")); if (vp->v_flag & VAGE) { TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); } else { TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); } freevnodes++; mtx_unlock(&vnode_free_list_mtx); vp->v_flag &= ~VAGE; vp->v_flag |= VFREE; splx(s); } /* * Opposite of vfree() - mark a vnode as in use. */ void vbusy(vp) struct vnode *vp; { int s; s = splbio(); mtx_lock(&vnode_free_list_mtx); KASSERT((vp->v_flag & VFREE) != 0, ("vnode not free")); TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); freevnodes--; mtx_unlock(&vnode_free_list_mtx); vp->v_flag &= ~(VFREE|VAGE); splx(s); } /* * Record a process's interest in events which might happen to * a vnode. Because poll uses the historic select-style interface * internally, this routine serves as both the ``check for any * pending events'' and the ``record my interest in future events'' * functions. (These are done together, while the lock is held, * to avoid race conditions.) */ int vn_pollrecord(vp, td, events) struct vnode *vp; struct thread *td; short events; { if (vp->v_pollinfo == NULL) v_addpollinfo(vp); mtx_lock(&vp->v_pollinfo->vpi_lock); if (vp->v_pollinfo->vpi_revents & events) { /* * This leaves events we are not interested * in available for the other process which * which presumably had requested them * (otherwise they would never have been * recorded). */ events &= vp->v_pollinfo->vpi_revents; vp->v_pollinfo->vpi_revents &= ~events; mtx_unlock(&vp->v_pollinfo->vpi_lock); return events; } vp->v_pollinfo->vpi_events |= events; selrecord(td, &vp->v_pollinfo->vpi_selinfo); mtx_unlock(&vp->v_pollinfo->vpi_lock); return 0; } /* * Note the occurrence of an event. If the VN_POLLEVENT macro is used, * it is possible for us to miss an event due to race conditions, but * that condition is expected to be rare, so for the moment it is the * preferred interface. */ void vn_pollevent(vp, events) struct vnode *vp; short events; { if (vp->v_pollinfo == NULL) v_addpollinfo(vp); mtx_lock(&vp->v_pollinfo->vpi_lock); if (vp->v_pollinfo->vpi_events & events) { /* * We clear vpi_events so that we don't * call selwakeup() twice if two events are * posted before the polling process(es) is * awakened. This also ensures that we take at * most one selwakeup() if the polling process * is no longer interested. However, it does * mean that only one event can be noticed at * a time. (Perhaps we should only clear those * event bits which we note?) XXX */ vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ vp->v_pollinfo->vpi_revents |= events; selwakeup(&vp->v_pollinfo->vpi_selinfo); } mtx_unlock(&vp->v_pollinfo->vpi_lock); } /* * Wake up anyone polling on vp because it is being revoked. * This depends on dead_poll() returning POLLHUP for correct * behavior. */ void vn_pollgone(vp) struct vnode *vp; { mtx_lock(&vp->v_pollinfo->vpi_lock); VN_KNOTE(vp, NOTE_REVOKE); if (vp->v_pollinfo->vpi_events) { vp->v_pollinfo->vpi_events = 0; selwakeup(&vp->v_pollinfo->vpi_selinfo); } mtx_unlock(&vp->v_pollinfo->vpi_lock); } /* * Routine to create and manage a filesystem syncer vnode. */ #define sync_close ((int (*)(struct vop_close_args *))nullop) static int sync_fsync(struct vop_fsync_args *); static int sync_inactive(struct vop_inactive_args *); static int sync_reclaim(struct vop_reclaim_args *); #define sync_lock ((int (*)(struct vop_lock_args *))vop_nolock) #define sync_unlock ((int (*)(struct vop_unlock_args *))vop_nounlock) static int sync_print(struct vop_print_args *); #define sync_islocked ((int(*)(struct vop_islocked_args *))vop_noislocked) static vop_t **sync_vnodeop_p; static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { { &vop_default_desc, (vop_t *) vop_eopnotsupp }, { &vop_close_desc, (vop_t *) sync_close }, /* close */ { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ { &vop_lock_desc, (vop_t *) sync_lock }, /* lock */ { &vop_unlock_desc, (vop_t *) sync_unlock }, /* unlock */ { &vop_print_desc, (vop_t *) sync_print }, /* print */ { &vop_islocked_desc, (vop_t *) sync_islocked }, /* islocked */ { NULL, NULL } }; static struct vnodeopv_desc sync_vnodeop_opv_desc = { &sync_vnodeop_p, sync_vnodeop_entries }; VNODEOP_SET(sync_vnodeop_opv_desc); /* * Create a new filesystem syncer vnode for the specified mount point. */ int vfs_allocate_syncvnode(mp) struct mount *mp; { struct vnode *vp; static long start, incr, next; int error; /* Allocate a new vnode */ if ((error = getnewvnode(VT_VFS, mp, sync_vnodeop_p, &vp)) != 0) { mp->mnt_syncer = NULL; return (error); } vp->v_type = VNON; /* * Place the vnode onto the syncer worklist. We attempt to * scatter them about on the list so that they will go off * at evenly distributed times even if all the filesystems * are mounted at once. */ next += incr; if (next == 0 || next > syncer_maxdelay) { start /= 2; incr /= 2; if (start == 0) { start = syncer_maxdelay / 2; incr = syncer_maxdelay; } next = start; } vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0); mp->mnt_syncer = vp; return (0); } /* * Do a lazy sync of the filesystem. */ static int sync_fsync(ap) struct vop_fsync_args /* { struct vnode *a_vp; struct ucred *a_cred; int a_waitfor; struct thread *a_td; } */ *ap; { struct vnode *syncvp = ap->a_vp; struct mount *mp = syncvp->v_mount; struct thread *td = ap->a_td; int asyncflag; /* * We only need to do something if this is a lazy evaluation. */ if (ap->a_waitfor != MNT_LAZY) return (0); /* * Move ourselves to the back of the sync list. */ vn_syncer_add_to_worklist(syncvp, syncdelay); /* * Walk the list of vnodes pushing all that are dirty and * not already on the sync list. */ mtx_lock(&mountlist_mtx); if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { mtx_unlock(&mountlist_mtx); return (0); } if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { vfs_unbusy(mp, td); return (0); } asyncflag = mp->mnt_flag & MNT_ASYNC; mp->mnt_flag &= ~MNT_ASYNC; vfs_msync(mp, MNT_NOWAIT); VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); if (asyncflag) mp->mnt_flag |= MNT_ASYNC; vn_finished_write(mp); vfs_unbusy(mp, td); return (0); } /* * The syncer vnode is no referenced. */ static int sync_inactive(ap) struct vop_inactive_args /* { struct vnode *a_vp; struct thread *a_td; } */ *ap; { vgone(ap->a_vp); return (0); } /* * The syncer vnode is no longer needed and is being decommissioned. * * Modifications to the worklist must be protected at splbio(). */ static int sync_reclaim(ap) struct vop_reclaim_args /* { struct vnode *a_vp; } */ *ap; { struct vnode *vp = ap->a_vp; int s; s = splbio(); vp->v_mount->mnt_syncer = NULL; if (vp->v_flag & VONWORKLST) { LIST_REMOVE(vp, v_synclist); vp->v_flag &= ~VONWORKLST; } splx(s); return (0); } /* * Print out a syncer vnode. */ static int sync_print(ap) struct vop_print_args /* { struct vnode *a_vp; } */ *ap; { struct vnode *vp = ap->a_vp; printf("syncer vnode"); if (vp->v_vnlock != NULL) lockmgr_printinfo(vp->v_vnlock); printf("\n"); return (0); } /* * extract the dev_t from a VCHR */ dev_t vn_todev(vp) struct vnode *vp; { if (vp->v_type != VCHR) return (NODEV); return (vp->v_rdev); } /* * Check if vnode represents a disk device */ int vn_isdisk(vp, errp) struct vnode *vp; int *errp; { struct cdevsw *cdevsw; if (vp->v_type != VCHR) { if (errp != NULL) *errp = ENOTBLK; return (0); } if (vp->v_rdev == NULL) { if (errp != NULL) *errp = ENXIO; return (0); } cdevsw = devsw(vp->v_rdev); if (cdevsw == NULL) { if (errp != NULL) *errp = ENXIO; return (0); } if (!(cdevsw->d_flags & D_DISK)) { if (errp != NULL) *errp = ENOTBLK; return (0); } if (errp != NULL) *errp = 0; return (1); } /* * Free data allocated by namei(); see namei(9) for details. */ void NDFREE(ndp, flags) struct nameidata *ndp; const uint flags; { if (!(flags & NDF_NO_FREE_PNBUF) && (ndp->ni_cnd.cn_flags & HASBUF)) { uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); ndp->ni_cnd.cn_flags &= ~HASBUF; } if (!(flags & NDF_NO_DVP_UNLOCK) && (ndp->ni_cnd.cn_flags & LOCKPARENT) && ndp->ni_dvp != ndp->ni_vp) VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); if (!(flags & NDF_NO_DVP_RELE) && (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { vrele(ndp->ni_dvp); ndp->ni_dvp = NULL; } if (!(flags & NDF_NO_VP_UNLOCK) && (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) { vrele(ndp->ni_vp); ndp->ni_vp = NULL; } if (!(flags & NDF_NO_STARTDIR_RELE) && (ndp->ni_cnd.cn_flags & SAVESTART)) { vrele(ndp->ni_startdir); ndp->ni_startdir = NULL; } } /* * Common filesystem object access control check routine. Accepts a * vnode's type, "mode", uid and gid, requested access mode, credentials, * and optional call-by-reference privused argument allowing vaccess() * to indicate to the caller whether privilege was used to satisfy the * request. Returns 0 on success, or an errno on failure. */ int vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) enum vtype type; mode_t file_mode; uid_t file_uid; gid_t file_gid; mode_t acc_mode; struct ucred *cred; int *privused; { mode_t dac_granted; #ifdef CAPABILITIES mode_t cap_granted; #endif /* * Look for a normal, non-privileged way to access the file/directory * as requested. If it exists, go with that. */ if (privused != NULL) *privused = 0; dac_granted = 0; /* Check the owner. */ if (cred->cr_uid == file_uid) { dac_granted |= VADMIN; if (file_mode & S_IXUSR) dac_granted |= VEXEC; if (file_mode & S_IRUSR) dac_granted |= VREAD; if (file_mode & S_IWUSR) dac_granted |= VWRITE; if ((acc_mode & dac_granted) == acc_mode) return (0); goto privcheck; } /* Otherwise, check the groups (first match) */ if (groupmember(file_gid, cred)) { if (file_mode & S_IXGRP) dac_granted |= VEXEC; if (file_mode & S_IRGRP) dac_granted |= VREAD; if (file_mode & S_IWGRP) dac_granted |= VWRITE; if ((acc_mode & dac_granted) == acc_mode) return (0); goto privcheck; } /* Otherwise, check everyone else. */ if (file_mode & S_IXOTH) dac_granted |= VEXEC; if (file_mode & S_IROTH) dac_granted |= VREAD; if (file_mode & S_IWOTH) dac_granted |= VWRITE; if ((acc_mode & dac_granted) == acc_mode) return (0); privcheck: if (!suser_cred(cred, PRISON_ROOT)) { /* XXX audit: privilege used */ if (privused != NULL) *privused = 1; return (0); } #ifdef CAPABILITIES /* * Build a capability mask to determine if the set of capabilities * satisfies the requirements when combined with the granted mask * from above. * For each capability, if the capability is required, bitwise * or the request type onto the cap_granted mask. */ cap_granted = 0; if (type == VDIR) { /* * For directories, use CAP_DAC_READ_SEARCH to satisfy * VEXEC requests, instead of CAP_DAC_EXECUTE. */ if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) cap_granted |= VEXEC; } else { if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT)) cap_granted |= VEXEC; } if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) cap_granted |= VREAD; if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT)) cap_granted |= VWRITE; if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT)) cap_granted |= VADMIN; if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { /* XXX audit: privilege used */ if (privused != NULL) *privused = 1; return (0); } #endif return ((acc_mode & VADMIN) ? EPERM : EACCES); }